Datasets:

sqres

/

rawpapers

like 0

Follow

sqres 2

synthetic_cpt

LLM-Neo_Parameter_Efficient_Knowledge_Distillation_for_Large_Language_Models.pdf

4 2 0 2 n u J 3 1 ] E S . s c [ 1 v 0 0 3 0 1 . 6 0 4 2 : v i X r a Large Language Models as Software Components: A Taxonomy for LLM-Integrated Applications Irene Weber Kempten University of Applied Sciences, Germany irene.weber@hs-kempten.de Abstract Large Language Models (LLMs) have become widely adopted recently. Research explores their use both as autonomous agents and as tools for software engineering. LLM-integrated applications, on the other hand, are software systems that leverage an LLM to perform tasks that would otherwise be impossible or require significant coding effort. While LLM-integrated application engineering is emerging as new discipline, its terminology, concepts and methods need to be established. This study provides a taxonomy for LLM- integrated applications, offering a framework for analyzing and describing these systems. It also demonstrates various ways to utilize LLMs in applications, as well as options for implementing such integrations. Following established methods, we analyze a sample of recent LLM-integrated applications to identify rel- evant dimensions. We evaluate the taxonomy by applying it to additional cases. This review shows that applications integrate LLMs in numerous ways for various purposes. Frequently, they comprise multiple LLM integrations, which we term “LLM components”. To gain a clear understanding of an application’s architecture, we examine each LLM component separately. We identify thirteen dimensions along which to characterize an LLM component, including the LLM skills leveraged, the format of the output, and more. LLM-integrated applications are described as combinations of their LLM components. We suggest a concise representation using feature vectors for visualization. The taxonomy is effective for describing LLM-integrated applications. It can contribute to theory building in the nascent field of LLM-integrated application engineering and aid in developing such systems. Researchers and practitioners explore numerous creative ways to leverage LLMs in applications. Though challenges persist, integrating LLMs may revolutionize the way software systems are built. Keywords: component large language model, LLM-integrated, taxonomy, copilot, architecture, AI agent, LLM 1. Introduction fields, such as medicine, law, marketing, education, human resources, etc. Large Language Models (LLMs) have significantly impacted various sectors of economy and society [47]. Due to their proficiency in text understanding, cre- ative work, communication, knowledge work, and code writing, they have been adopted in numerous Public discussions often focus on the ethical aspects and societal consequences of these systems [36, 39]. Meanwhile, research investigates Artificial General Intelligences and autonomous AI agents that can use services, data sources, and other tools, and collabo- rate to solve complex tasks [11, 62, 57, 21]. In addi- tion, LLMs offer many opportunities to enhance soft- ware systems. They enable natural language interac- tion [59], automate complex tasks [19], and provide supportive collaboration, as seen with recent LLM- based assistant products often branded as “copilots” 1. This paper addresses the potential of LLMs for soft- ware development by integrating their capabilities as components into software systems. This contrasts with current software engineering research, which views LLMs as tools for software development rather than as software components [14, 22], and with the considerable body of research examining LLMs as au- tonomous agents within multiagent systems [21]. Software systems that invoke an LLM and process its output are referred to as “LLM-integrated appli- cations”, “LLM-integrated systems”, “LLM-based ap- plications”, etc. [32, 13, 57]. LLMs are versatile, mul- tipurpose tools capable of providing functionalities that would otherwise be unfeasible or require sub- stantial development efforts [15, 24]. By significantly expediting system development, they have the poten- tial to revolutionize not only the way users interact with technology, but also the fundamental processes of software development. LLM-integrated applications engineering is emerging as a research field. E.g., [10] proposes LLM Sys- tems Engineering (LLM-SE) as a novel discipline, and [44, 8, 7] discuss experiences and challenges that de- velopers of such systems encounter in practice. This study develops a taxonomy that provides a structured framework for categorizing and analyzing LLM-integrated applications across various domains. To develop and evaluate the taxonomy, we collected a sample of LLM-integrated applications, concentrat- ing on technical and industrial domains. These ap- plications showcase a broad range of opportunities to leverage LLMs, often integrating LLMs in mul- tiple ways for distinct purposes. In developing the taxonomy, we found that examining each of these in- tegrations, termed “LLM components”, separately is crucial for a clear understanding of an application’s architecture. The taxonomy adopts an original architectural per- spective, focusing on how the application interacts with the LLM while abstracting from the specifics of application domains. For researchers, the taxon- omy contributes to shape a common understanding and terminology, thus aiding theory building in this emerging domain [29, 50, 18]. For practitioners, the taxonomy provides inspiration for potential uses of LLMs in applications, presents design options, and helps identify challenges and approaches to address them. Objectives. In this study, a taxonomy is understood as a set of dimensions divided into characteristics. The objective is to identify dimensions that are useful for categorizing the integration of LLMs in applica- tions from an architectural perspective. To be most effective, the taxonomy should be easy to understand and apply, yet distinctive enough to uncover the es- sential aspects. Additionally, we aim to develop a visual representation tailored to the taxonomy’s in- tended purposes. Overview. The following section 2 provides back- ground on LLMs and introduces relevant concepts. Section 3 presents an overview of related work. The study design adheres to a Design Science Research approach [46]. We apply established methods for tax- onomy design [42, 48] as described in Section 4. This section also presents the sample of LLM-integrated applications used for this study. The developed tax- onomy is presented, demonstrated and formally eval- uated in section 5. In section 6, we discuss its usabil- ity and usefulness. Section 7 summarizes the contri- butions, addresses limitations, and concludes. 2. Large Language Models 2.1. Background 1E.g., https://docs.github.com/en/copilot, https://copilot.cloud.microsoft/en-us/copilot-excel, https://www.salesforce.com/einsteincopilot State-of-the-art LLMs such as GPT-3.5, GPT-4, Llama, PALM2, etc., are artificial neural networks i.e., very simple processing consisting of neurons, 2 units, that are organized in layers and connected by weighted links. Training a neural network means adapting these weights such that the neural network shows a certain desired behavior. Specifically, an LLM is trained to predict the likelihoods of pieces of text termed, tokens, to occur as continuations of a given text presented as input to the LLM. This in- put is referred to as prompt. The prompt combined with the produced output constitutes the context of an LLM. It may comprise more than 100k tokens in state-of-the-art LLMs2. Still, its length is limited and determines the maximum size of prompts and outputs that an LLM is capable of processing and generating at a time. Training of an LLM optimizes its parameters such that its computed likelihoods align with real text ex- amples. The training data is a vast body of text snip- pets extracted, processed, and curated from sources such as Wikipedia, Github code repositories, common websites, books, or news archives. An LLM trained on massive examples is termed a foundation model or pre-trained model. During training, an LLM not only learns to produce correct language but also ab- sorbs and stores information and factual knowledge. However, it is well known that LLMs frequently pick up biases, leading to ethical problems. They may also produce factually incorrect outputs that sound plausible and convincing, termed hallucinations. Recent findings show that LLMs can be applied to a wide range of tasks by appropriately formulating prompts. Different prompt patterns succeed in dif- ferent tasks. Basic approaches rely on instructing the LLM to solve a task described or explained in the prompt. In few-shot prompting (also known as few-shot learning), the prompt is augmented with ex- ample input-output pairs illustrating how to solve the task, e.g., the requested output format. The number of examples can vary. Prompting with one example is called one-shot prompting, while prompting without any examples is called zero-shot prompting. One-shot and few-shot prompting fall under the broader cat- egory of in-context learning. Prompt patterns such 2https://platform.openai.com/docs/models as chain-of-thought and thinking-aloud aim to elicit advanced reasoning capabilities from LLMs. As effective prompts are crucial for unlocking the di- verse capabilities of an LLM, the discipline of prompt engineering is evolving, focusing on the systematic design and management of prompts [66, 9, 53, 31]. 2.2. Definitions Invoking an LLM results in an input-processing- output sequence: Upon receiving a prompt, the LLM processes it and generates an output. We refer to an individual sequence of input-processing-output per- formed by the LLM as LLM invocation, and define an LLM-integrated application as a system in which the software generates the prompt for the LLM and processes its output. The concept of an application is broad, encompassing service-oriented architectures and systems with components loosely coupled via API calls. Given an LLM’s versatility, an application can uti- lize it for different tasks, each demanding a specific approach to create the prompt and handle the re- sult. This paper defines a particular software compo- nent that accomplishes this as an LLM-based software component or, simply, LLM component. An LLM- integrated application can comprise several LLM components. The study develops a taxonomy for LLM components. LLM-integrated applications are described as combinations of their LLM components. 3. Related Work With the recent progress in generative AI and LLMs, the interest in these techniques has increased, and numerous surveys have been published, providing an extensive overview of technical aspects of LLMs [72], reviewing LLMs as tools for software engineering [22], and discussing the technical challenges of applying LLMs across various fields [25]. Further studies ad- dress the regulatory and ethical aspects of Genera- tive AI and ChatGPT, with a particular focus on AI-human collaboration [41], and Augmented Lan- guage Models (ALMs), which are LLMs that enhance 3 their capabilities by querying tools such as APIs, databases, and web search engines [38]. Taxomonies related to LLMs include a taxonomy for prompts designed to solve complex tasks [49] and a taxonomy of methods for cost-effectively invoking a remote LLM [60]. A comparative analysis of stud- ies on applications of ChatGPT is provided by [27], whereas LLMs are compared based on their applica- tion domains and the tasks they solve in [20]. Most closely related to the taxonomy developed here is a taxonomy for LLM-powered multiagent architectures [21] which focuses on autonomous agents with less technical detail. Taxonomies of applications of AI in enterprises [48] and applications of generative AI, in- cluding but not limited to LLMs [52], are developed using methods similar to those in our study. Several taxonomies in the field of conversational agents and task-oriented dialog (TOD) systems ad- dress system architecture [1, 40, 12, 3]. However, they omit detailed coverage of the integration of generative language models. 4. Methods We constructed the taxonomy following established guidelines [42, 48, 29], drawing from a sample of LLM-integrated applications. These applications are detailed in section 4.1. 4.1. Development Taxonomy. We derived an initial taxonomy from the standard architecture of conversational assistants de- scribed in [3], guided by the idea that conversational assistants are essentially “chatbots with tools”, i.e., language-operated user interfaces that interact with external systems. This approach proved unsuccessful. The second version was based on the classical three- tier software architecture, and then extended over several development cycles. By repeatedly apply- ing the evolving taxonomy to the example instances, we identified dimensions and characteristics using an “empirical-to-conceptual” approach. When new di- mensions emerged, additional characteristics were de- rived in a “conceptual-to-empirical” manner. After five major refinement cycles, the set of dimensions and characteristics solidified. In the subsequent eval- uation phase, we applied the taxonomy to a new set of example instances that were not considered while constructing the taxonomy. As the dimensions and characteristics remained stable, the taxonomy was considered complete. In the final phase, we refined the wording and visual format of the taxonomy. Visualization. Developing a taxonomy involves cre- ating a representation that effectively supports its intended purpose [29]. Taxonomies can be repre- sented in various formats, with morphological boxes [54, 55] or radar charts [21] being well-established approaches. We evaluated morphological boxes, be- cause they effectively position categorized instances within the design space. However, we found that they make it difficult to perceive a group of categorized in- stances as a whole since they occupy a large display area. This drawback is significant for our purposes, as LLM-integrated applications often comprise mul- tiple LLM components. Therefore, we developed a more condensed visualization of the taxonomy based on feature vectors. Example instances. We searched for instances of LLM-integrated applications for taxonomy develop- ment that should meet the following criteria: • The application aims for real-world use rather than focusing on research only (such as testbeds for experiments or proofs-of-concept). It demon- strates efforts towards practical usability and ad- dresses challenges encountered in real-world sce- narios. • The application’s architecture, particularly its LLM components, is described in sufficient de- tail for analysis. • The sample of instances covers a diverse range of architectures. • The example instances are situated within indus- trial or technical domains, as we aim to focus on LLM-integrated applications beyond well-known fields like law, medicine, marketing, human re- sources, and education. 4 The search revealed a predominance of theoretical re- search on LLM-integrated applications while papers focusing on practically applied systems were scarce. Searching non-scientific websites uncovered commer- cially advertised AI-powered applications, but their internal workings were typically undisclosed, and reli- able evaluations were lacking. Furthermore, the het- erogeneous terminology and concepts in this emerg- literature ing field make a comprehensive formal search unfeasible. Instead, by repeatedly search- ing Google Scholar and non-scientific websites using terms “LLM-integrated applications”, “LLM-powered applications”, “LLM-enhanced system”, “LLM” and “tools”, along similar variants, we selected six suitable instances. Some of them integrate LLMs in multiple ways, totaling eleven distinct LLM components. For a thorough evaluation, we selected new instances using relaxed criteria, including those intended for research. Additionally, we included a real-world ex- ample lacking explicit documentation to broaden the diversity of our sample and assess the taxonomy’s coverage. Within the five selected instances, we iden- tified ten LLM components. 4.2. Sample of LLM-integrated applications Table 1 gives an overview of the sample. Names of ap- plications and LLM components are uniformly writ- ten as one CamelCase word and typeset in small caps, deviating from the format chosen by the respective authors. LowCode. LowCode is a web-based application consisting of a prompt-definition section and a di- alogue section. The prompt-definition section sup- ports the design of prompts for complex tasks, such as composing extensive essays, writing resumes for job applications or acting as a hotel service chatbot [5]. In the dialogue section, users converse with an LLM to complete the complex task based on the de- fined prompt. LowCode comprises two LLM components termed Planning and Executing. Planning operates in the prompt-definition section, where a user roughly describes a complex task, and Planning designs a workflow for solving it. The prompt-definition section offers a low-code development environment where the LLM-generated workflow is visualized as a graphi- cal flowchart, allowing a user to edit and adjust the logic of the flow and the contents of its steps. For instance, in essay-writing scenarios, this involves in- serting additional sections, rearranging sections, and refining the contents of sections. Once approved by the user, LowCode translates the modified work- flow back into natural language and incorporates it into a prompt for Executing. In the dialogue sec- tion, users converse in interactive, multi-turn dia- logues with Executing. As defined in the prompt, it acts as an assistant for tasks such as writing an essay or resume, or as a hotel service chatbot. While the idea of the LLM planning a workflow might suggest using the LLM for application control, LowCode Planning actually serves as a prompt generator that supports developing prompts for complex tasks. Honeycomb. Honeycomb is an observability plat- form collecting data from software applications in distributed environments for monitoring. Users define queries to retrieve information about the observed software systems through Honeycomb’s Query Builder UI. The recently added LLM-based QueryAssistant allows users to articulate inquiries in plain English, such as “slow endpoints by status code” or “which service has the highest latency?” The QueryAssistant converts these into queries in Honeycomb’s format, which users can execute and manually refine [7, 8]. MyCrunchGpt. MyCrunchGpt acts as an ex- pert system within the engineering domain, specif- ically for airfoil design and calculations in fluid me- chanics. These tasks require complex workflows com- prising several steps such as preparing data, param- eterizing tools, and evaluating results, using vari- ous software systems and tools. The aim of My- CrunchGpt is to facilitate the definition of these workflows and automate their execution [28]. MyCrunchGpt offers a web interface featuring a dialogue window for inputting commands in plain English, along with separate windows displaying the 5 Table 1: Example instances selected for development (top 6) and evaluation (bottom 5) Application References LLM components Honeycomb QueryAssistant [7, 8] Planning, Executing LowCode [5],[35] DesignAssistant, SettingsEditor, DomainExpert [28] MyCrunchGpt Manager, Operator MatrixProduction [69] TaskPlanning [37] WorkplaceRobot TaskExecutor, MemoryGenerator [64] AutoDroid ActionPlanning, ScenarioFeedback [51] ProgPrompt QuestionAnswering [26] FactoryAssistants DstPrompter, PolicyPrompter [71] SgpTod Reporting [70] TruckPlatoon ActionExecutor, Advisor, IntentDetector, Explainer [16, 44] ExcelCopilot output and results of software tools invoked by My- CrunchGpt in the backend. MyCrunchGpt relies on predefined workflows, not supporting deviations or cycles. By appending a specific instruction to the dialogue history in the prompt for each step of the workflow, it uses the LLM as a smart parser to ex- tract parameters for APIs and backend tools from user input. APIs and tools are called in the prede- fined order [28, p. 56]. MyCrunchGpt is still in development. The paper [28] explains the domain as well as the integration of the LLM, but does not fully detail the implementa- tion of the latter. Still, MyCrunchGpt illustrates innovative applications of an LLM in a technical do- main. We categorize three LLM components solving tasks within MyCrunchGpt: a DesignAssistant guiding users through workflows and requesting pa- rameters for function and API calls; a SettingsEd- itor updating a JSON file with settings for a back- end software tool; and a DomainExpert which helps evaluating results by comparing them to related re- sults, e.g., existing airfoil designs, which it derives from its trained knowledge. MatrixProduction. MatrixProduction em- ploys an LLM for controlling a matrix production system [69]. While in a classical line production setup, workstations are arranged linearly and the manufacturing steps follow a fixed sequence, matrix production is oriented towards greater flexibility. transport vehicles Autonomous carry materials and intermediate products to workstations, termed automation modules, each offering a spectrum of manufacturing skills that it can contribute to the production process. Compared to line production, matrix production is highly adaptable and can manufacture a variety of personalized products with full automation. This requires intelligent production management to (a) create workplans that orchestrate and schedule the automation modules’ skills, and (b) program the involved automation modules such that they execute the required processing steps. MatrixProduction incorporates two LLM compo- nents: Manager creates workplans as sequences of skills (a), while Operator generates programs for the involved automation modules (b). MatrixProduction prompts Manager and Op- erator to provide textual explanations in addition to the required sequences of skills or automation module programs. The LLM output is processed by a parser before being used to control the physi- cal systems. Manager relies on built-in production- specific knowledge of the LLM such as “a hole is pro- duced by drilling”. Noteworthy in this approach is its tight integra- tion into the system landscape of Industry 4.0. The few-shot Manager and Operator prompts are generated automatically using Asset Adminis- tration Shells, which are standardized, technology- 6 independent data repositories storing digital twins of manufacturing assets for use in Industry 4.0 [2]. WorkplaceRobot. An experimental robot system is enhanced with LLM-based task planning in [37]. The robot operates in a workplace environment fea- turing a desk and several objects. It has previously been trained to execute basic operations expressed in natural language such as “open the drawer” or “take the pink object and place it in the drawer”. LLM-based task planning enables the robot to per- form more complex orders like “tidy up the work area and turn off all the lights”. To this end, an LLM is prompted to generate a sequence of basic operations that accomplish the complex order. Although the robot expects operations phrased in language, the LLM is prompted with a natural Python coding task. For instance, the basic opera- tion “turn on the green light” corresponds to a Python command push_button(’green’). The prompt for the LLM includes several examples each consisting of a description of an environment state, a complex order formatted as a comment, and a sequence of Python robot commands that accomplish the com- plex order. When invoking the LLM to generate the Python program for a new order, the prompt is aug- mented with a description of the environment’s cur- rent state and the new order as a comment. The Python code produced by the LLM is trans- lated back to a sequence of basic operations in nat- ural language. When the robot executes these oper- ations, there is no feedback about successful comple- tion. Rather, the system assumes that all basic op- erations require a fixed number of timesteps to com- plete. AutoDroid. The goal of mobile task automation is hands-free user interaction for smartphones through voice commands. AutoDroid is a voice control sys- tem for smartphones that can automatically execute complex orders such as “remind me to do laundry on May 11th” or “delete the last photo I took” [64, 65]. as “scroll down, then press button x” in the calen- dar app. AutoDroid employs an LLM component TaskExecutor to plan these sequences of opera- tions. The challenge is that the next operation to ex- ecute depends on the current state of the Android app which continuously changes as the app is operated. AutoDroid solves this by invoking the TaskEx- ecutor repeatedly after each app operation with the prompt comprising the updated state of the Graph- ical User Interface (GUI) along with the user’s com- plex order. Before executing irrevocable operations, such as per- manently deleting data or calling a contact, Auto- Droid prompts the user to confirm or adjust the op- eration. TaskExecutor is instructed to include a “confirmation needed” hint in its output for such op- erations. The prompt for TaskExecutor comprises an ex- tract from a knowledge base which is built automati- cally in an offline learning phase as follows: In a first step, a “UI Automator” (which is not an LLM com- ponent) automatically and randomly operates the GUI elements of an Android app to generate a UI Transition Graph (UTG). The UTG has GUI states as nodes and the possible transitions between GUI states as edges. As next steps, AutoDroid invokes two LLM components referred to as MemoryGen- erators to analyze the UTG. The first MemoryGenerator is prompted repeat- edly for each GUI state in the UTG. Its task is to explain the functionality of the GUI elements. Be- sides instructions and examples of the table format desired as output, its prompt includes an HTML rep- resentation of the GUI state, the GUI actions preced- ing this state, and the GUI element operated next. Its output consists of tuples explaining the function- ality of a GUI element by naming the derived func- tionality (e.g., “delete all the events in the calendar app”) and the GUI states and GUI element actions in- volved. Similarly, the second MemoryGenerator is prompted to output a table listing GUI states and explanations of their functions. These tables consti- tute AutoDroid’s knowledge base. Such complex orders are fulfilled by performing se- quences of basic operations in an Android app, such ProgPrompt. ProgPrompt [51] is an approach to to LLM-based robot task planning similar 7 Its robot is controlled by WorkplaceRobot. Python code and works in a real and a simulated household environment. ProgPrompt comprises two LLM components. Ac- tionPlanning generates Python scripts for tasks such as “microwave salmon” using basic opera- tions like grab(’salmon’), open(’microwave’), and putin(’salmon’, ’microwave’), notably with- out considering the current state of the environment. To establish a feedback loop with the environment, ActionPlanning adds assert statements. These statements verify the preconditions of basic opera- tions and trigger remedial actions when preconditions are not met. For instance, a script for “microwave salmon” comprises the following code fragment: if assert(’microwave’ is ’opened’) else: open(’microwave’) putin(’salmon’, ’microwave’) When operating in the simulated environment, ProgPrompt can verify an assert statement through its second LLM component, Scenario- Feedback. Prompted with the current state of the environment and the assert statement, Scenario- Feedback evaluates it and outputs True or False. FactoryAssistants. FactoryAssistants advise workers on troubleshooting production line issues in two manufacturing domains: detergent production and textile production [26]. The assistants leverage domain knowledge from FAQs and documented prob- lem cases to answer user queries. The required do- main knowledge is provided as a part of the prompt. SgpTod. SgpTod employs an LLM to implement a chatbot, specifically, a task-oriented dialogue (TOD) system [71]. TOD systems are also known as conver- sational assistants. In contrast to open-domain dia- logue (ODD) systems, which engage users in goalless conversations, they are designed for assisting users in specific tasks. In general, TOD systems require the following components [3]: Natural Language Understanding (NLU), analyzing the user’s input to classify intents and extract entities; Dialogue Management (DM) for deciding on a system action that is appropriate in a given dialogue state (e.g., ask for more informa- tion or invoke a hotel booking service); and Natu- ral Language Generation (NLG) for producing a re- sponse that the TOD system can present to the user. Intent classification, also known as intent detection, matches free-text user input to one of several tasks a TOD system can perform (e.g., book a hotel). Entity extraction isolates situational values, called entities, from the user input (e.g., the town and the date of the hotel booking). The TOD system may require several dialogue turns to elicit all necessary entities from the user. In TOD research, the system’s in- ternal representation of the user’s intentions and the entity values is commonly referred to as its “belief state”. For example, in the restaurant search domain, the belief state may include attribute-value pairs like cuisine:Indian and pricerange:medium. SgpTod is a multi-domain TOD system, concur- rently handling multiple task domains found in stan- dard TOD evaluation datasets, such as recommend- ing restaurants or finding taxis. Similar to other ex- perimental TOD systems [23], SgpTod accesses a database that stores information from the task do- mains, such as available hotels and restaurants. SgpTod comprises two LLM components, called DstPrompter and PolicyPrompter, that are both invoked in every dialogue turn between SgpTod and the user. The DstPrompter handles the NLU aspect, analyzing the user’s input and populating the system’s belief state. It outputs is an SQL query suited to extract the database entries that match the current belief state. Upon retrieving the database en- tries, SgpTod invokes its PolicyPrompter which covers both DM and NLG. Prompted with the dia- logue history and the database entries retrieved, it produces a two-part output: a natural language re- sponse for NLG and a system action for DM. TruckPlatoon. The concept of truck platooning means that trucks travel closely together for bet- ter fuel efficiency and traffic flow. TruckPla- toon comprises an algorithmic control loop which autonomously maintains a consistent distance be- tween trucks. It invokes an LLM to generate natural- language reports on the platoon’s performance and 8 stability from measurements tracked by the control algorithm, providing easily understandable informa- tion for engineers involved in monitoring and opti- mizing the truck platooning system. ExcelCopilot. ExcelCopilot is an example of a recent trend where software companies integrate LLM-based assistants, often termed “copilots”, into their products [44]. These copilots not only provide textual guidance but also perform actions within the software environment, constituting a distinctive type of LLM-integrated application. We chose Excel- Copilot as an example for evaluating our taxonomy. Since its implementation is undisclosed, we infer its architecture from indirect sources, including a screen- cast and a report on insights and experiences from copilot developers [16, 44]. This inferred architecture may deviate from the actual implementation. ExcelCopilot is accessible in a task bar along- side the Excel worksheet. It features buttons with context-dependent suggestions of actions and a text box for users to type in commands in natural lan- guage. ExcelCopilot only works with data tables, so its initial suggestion is to convert the active work- sheet’s data into a data table. Copilot functions ac- tivate when a data table or part of it is selected. It then presents buttons for four top-level tasks: “add formula columns”, “highlight”, “sort and filter”, and “analyze”. The “analyze” button triggers the copilot to display more buttons, e.g., one that generates a pivot chart from the selected data. ExcelCopilot can also add a formula column to the data table and explain the formula in plain language. When a user inputs a free-text command, Excel- Copilot may communicate its inability to fulfill it. This constantly occurs with commands requiring multiple steps, indicating that ExcelCopilot lacks a planning LLM component as seen in, for example, MatrixProduction. This observation, along with its mention in [44], suggests that ExcelCopilot em- ploys an intent detection-skill routing architecture. This architecture includes an LLM component that maps free-text user commands to potential intents and then delegates to other LLM components tasked with generating actions to fulfill those intents. Ac- cordingly, ExcelCopilot comprises several types of LLM components: • Several distinct Action Executors generate code for specific application actions, such as cre- ating a pivot table, designing a worksheet for- mula, inserting a diagram, and so on. • An Advisor suggests meaningful next actions. Its outputs serve to derive button captions and prompts for ActionExecutors. • When a user inputs a free-text command, the IntentDetector is invoked to determine and trigger a suitable ActionExecutor. The In- tentDetector communicates its actions to users and informs them when it cannot devise a suitable action. • The Explainer generates natural language ex- planations of formulae designed by ExcelCopi- lot. It is unclear whether under the hood, the ActionExecutor is generating both the for- mula and the explanation, or if two separate LLM components are being invoked. We assume the latter, i.e., that a separate Explainer LLM component exists. While users interact repeatedly with ExcelCopi- lot, each interaction adheres to a single-turn pat- tern, with the user providing a command and Ex- celCopilot executing it [44]. 5. A Taxonomy for LLM Components and LLM-Integrated Applications When developing the taxonomy, it emerged that an- alyzing an LLM-integrated application should begin with identifying and describing its distinct LLM com- ponents. Analyzing each LLM component separately helps capture details and provides a clear understand- ing of how the application utilizes LLM capabili- ties. The LLM-integrated application can then be described as a combination of the LLM components it employs. 9 Function Meta Invocation Table 2: Dimensions and characteristics of the taxonomy. Codes of characteristics are printed in uppercase. “Meta” means “metadimension”. “MuEx” means “mutual exclusiveness”. Dimension Interaction Frequency Logic UI Data Instruction State Task Check Skills Format Revision Consumer Characteristics App, Command, Dialog Single, Iterative cAlculate, Control none, Input, Output, Both none, Read, Write, Both none, User, LLM, Program none, User, LLM, Program none, User, LLM, Program none, User, LLM, Program reWrite, Create, conVerse, Inform, Reason, Plan FreeText, Item, Code, Structure none, User, LLM, Program User, LLM, Program, Engine MuEx enforced yes yes yes yes enforced enforced yes enforced no no enforced enforced Prompt Output 5.1. Overview and demonstration The taxonomy identifies 13 dimensions for LLM com- ponents, grouped into five metadimensions as shown in table 2. It comprises both dimensions with gen- uinely mutually exclusive characteristics and those with non-exclusive characteristics. For dimensions related to the technical integration of LLMs within applications, mutual exclusiveness is enforced. Given the open nature of software architecture, the inte- gration of LLMs allows for significant diversity. In practice, LLM components may show multiple char- acteristics within these dimensions. Nonetheless, the taxonomy requires categorizing each component with a predominant characteristic, enforcing a necessary level of abstraction to effectively organize and struc- ture the domain. We applied the taxonomy to categorize each of the example instances described in section 4.2. The re- sults are depicted in figure 1. The dimensions and their characteristics are detailed and illustrated with examples in section 5.2. The taxonomy visualizes an LLM component by a feature vector comprising binary as well as multi- valued features. Non-mutually exclusive dimensions are represented by a set of binary features. The re- maining dimensions are encoded as n-valued features where n denotes the number of characteristics. For compactness, we use one-letter codes of the charac- teristics as feature values in the visualizations. In table 2, these codes are printed in upper case in the respective characteristic’s name. A feature vector representing an LLM component is visualized in one line. For dimensions with non- mutually exclusive characteristics, all possible codes are listed, with the applicable ones marked. The re- maining dimensions are represented by the code of the applicable characteristic, with the characteris- tic none shown as an empty cell. We shade feature values with different tones to support visual percep- tion. LLM components within the same application are grouped together, visualizing an LLM-integrating application in a tabular format. 5.2. Dimensions and characteristics 5.2.1. Invocation dimensions Two Invocation dimensions address the way the LLM is invoked within the application. Interaction describes how the user interacts with the LLM with three characteristics: App: Users never converse with the LLM directly in natural language, rather the application invokes the LLM automatically. E.g., users do not interact 10 Invocation Function Prompt (cid:125)(cid:124) (cid:123) (cid:122) (cid:125)(cid:124) (cid:123) (cid:125)(cid:124) (cid:123) (cid:122) Skills (cid:125)(cid:124) Out. Format Output (cid:123) (cid:122) (cid:125)(cid:124) (cid:123) (cid:122) (cid:125)(cid:124) (cid:123) (cid:122) n o i t c a r e t n I C C D Honeycomb QueryAssistant LowCode Planning LowCode Executing MyGrunchGpt DesignAssistant D C MyGrunchGpt SettingsEditor C MyGrunchGpt DomainExpert MatrixProduction Manager MatrixProduction Operator WorkplaceRobot AutoDroid Executor AutoDroid MemoryGenerator2 C A C C A C ProgPrompt ActionPlanning ProgPrompt ScenarioFeedback A FactoryAssistant SgpTod DstPrompter SgpTod PolicyPrompter TruckPlatoon D D A A ExcelCopilot ActionExecutor∗ A A ExcelCopilot Advisor C ExcelCopilot IntentDetector A ExcelCopilot Explainer y c n e u q e r F S S I I S S S S S I I S I S S S S S S S S (cid:122) n o i t c u r t s n I a t a D I U c i g o L A e t a t S k s a T k c e h C e t i r W e r e t a e r C e s r e V n o c m r o f n I n o s a e R A A B A B A A I I I I C C C C A C C A R P P U P P U P L U P P U P P P P P P P P U P P L P P U I C V I V W I I P L U P P P P P U P P L P P U W V V A I R P P U P P P C O A O P P P W A A C A P P L P P P P P U P P P t x e T e e r F m e t I n a l P P P F F P F P F P P P F F F P F F F R R R R R R R I I I I e d o C C C C C C C e r u t c u r t S n o i s i v e R r e m u s n o C P E S U L U S S S S S S S E E U L E E E L E E U E P U E P P U Figure 1: Categorized example instances. See table 2 for a legend. ∗, 2: multiple LLM components. directly with ExcelCopilot ActionExecutor or with MatrixProduction Operator. Command : Users input single natural language commands. E.g., users interact with AutoDroid TaskExecutor through single natural language commands. Dialog: Users engage in multi-turn dialogues with the LLM component to achieve a use goal. E.g., users repeatedly prompt LowCode Executing or My- CrunchGpt DesignAssistant in multi-turn dia- logues to obtain an essay or an airfoil design, respec- tively. Frequency addresses how often the application in- vokes a specific LLM component to fulfill a goal: Single: A single invocation of an LLM component is sufficient to produce the result. E.g., in My- CrunchGpt, the application internally invokes dis- tinct LLM components once for each user input by injecting varying prompt instructions. Iterative: The LLM component is invoked repeatedly to produce the result. E.g., AutoDroid TaskEx- 11 ecutor is invoked multiple times to fulfill a com- mand with an updated environment description in the State prompt; LowCode Executing is repeat- edly prompted by the user to achieve the use goal while the application updates the dialogue history. 5.2.2. Function dimensions The Function dimensions are derived from the classi- cal three-tier software architecture model which seg- regates an application into three distinct layers: pre- sentation, logic and data [17]. The presentation layer implements the UI. On the input side, it allows users to enter data and commands that control the appli- cation. On the output side, it presents information and provides feedback on the execution of commands. The logic layer holds the code that directly realizes the core objectives and processes of an application such as processing data, performing calculations, and making decisions. The data layer of an application manages the reading and writing of data from and to persistent data storage. Due to its versatility, an LLM component can simultaneously implement func- tionality for all three layers. The taxonomy addresses this with three Function dimensions. UI indicates whether an LLM component contributes significantly to the user interface of an application, avoiding the need to implement graphical UI controls or display elements: none: No UI functionality is realized by the LLM. E.g., in ExcelCopilot, the LLM does not replace any UI elements. Input: is (partially) implemented by the LLM. E.g., in MatrixProduction Manager, users input their order in natural language, obviating a product configuration GUI. Output: Output UI is (partially) implemented by the LLM. E.g., in TruckPlatoon, the output gener- ated by the LLM component can replace a data cock- pit with gauges and other visuals displaying numeri- cal data. Input and output UI are (partially) imple- Both: mented by the LLM. E.g., in MyCrunchGpt, the DesignAssistant provides a convenient conversa- interface for parameterization of APIs and tional Input UI tools and feedback on missing values, which other- wise might require a complex GUI. Logic indicates whether the LLM component deter- mines the control flow of the application. It discerns two characteristics: cAlculate: The output does not significantly impact the control flow of the application, i.e., the output is processed like data. E.g., MyCrunchGpt Set- tingsEditor modifies a JSON file, replacing a pro- grammed function; MyCrunchGpt DesignAssis- tant asks the user for parameters, but the sequence of calling APIs and tools follows a predefined work- flow; the workflow computed by LowCode Plan- ning is displayed without influencing the applica- tion’s control flow. Control : The output of the LLM is used for con- trolling the application. E.g., the plans generated by MatrixProduction Manager serve to sched- ule and activate production modules; the actions pro- posed by AutoDroid TaskExecutor are actually executed and determine how the control flow of the app proceeds. Since an LLM invocation always computes a result, cAlculate is interpreted as “calculate only”, making cAlculate and Control mutually exclusive. Data addresses whether the LLM contributes to read- ing or writing persistent data: none: The LLM does not contribute to reading or writing persistent data. This characteristic applies to most sample instances. Read : The LLM is applied for reading from persistent data store. E.g., SgpTod DstPrompter generates SQL queries which the application executes; Honey- comb QueryAssistant devises analytical database queries. Write and Both: No LLM component among the samples generates database queries for creating or updating persistent data. 5.2.3. Prompt-related dimensions Integrating an LLM into an application poses spe- cific requirements for prompts, such as the need for prompts to reliably elicit output in the requested 12 form [68]. While a broad range of prompt patterns have been identified and investigated [66], there is still a lack of research on successful prompt pat- terns specifically for LLM-integrated applications, on which this taxonomy could build. Developing prompt taxonomies is a challenging research endeavor in itself [49] and is beyond the scope of this research. There- fore, the taxonomy does not define a dimension with specific prompt patterns as characteristics, but rather focuses on how the application generates the prompt for an LLM component from a technical perspective. Prompts generally consist of several parts with dis- tinct purposes, generated by different mechanisms. Although many authors explore the concepts, a com- mon terminology has yet to be established. This is illustrated in table 3, showing terms from an ad-hoc selection of recent papers addressing prompt gener- In the table, italics indicate ation in applications. that the authors refrain from introducing an abstract term and instead use a domain-specific description. The term “examples” indicates a one-shot or few-shot prompt pattern. The terms that are adopted for the taxonomy are underlined. The taxonomy distinguishes three prompt parts re- ferred to as Prompt Instruction, Prompt State, and Prompt Task. These parts can occur in any order, potentially interleaved, and some parts may be ab- sent. • Instruction is the part of a prompt that outlines how to solve the task. Defined during LLM com- ponent development, it remains static through- out an application’s lifespan. • State is the situation-dependent part of the prompt that is created dynamically every time the LLM is invoked. The taxonomy opts for the term State instead of “context” in order to avoid confusion with the “LLM context” as explained in section 2. The State may include the current dialogue history, an extract of a knowledge base needed specifically for the current LLM invoca- tion, or a state or scene description, etc. • Task is the part of the prompt conveying the task to solve in a specific invocation. Prompt Instruction, State and Task describe the ori- gins of the prompt parts by uniform characteristics: none: The prompt part is not present. E.g., Prog- Prompt ActionPlanning has no State prompt, nor does LowCode Planning (except the dialogue history when planning a subprocess). Instruction and Task prompt parts are present in all sample in- stances. User : The user phrases the prompt part. E.g., the Task for ExcelCopilot IntentDetector or for LowCode Planning is phrased by the user. There are no sample instances where the user provides the Instruction or State prompt parts. LLM : The prompt part is generated by an LLM. E.g., LowCode Planning generates the State for Low- Code Executing and ExcelCopilot IntentDe- tector generates the Task for ExcelCopilot Ac- tionExecutors. Program: Application code generates the prompt part. E.g., AutoDroid programmatically generates the State and the Task parts for its MemoryGen- erators in the knowledge base building phase. The Prompt Instruction dimension is always gener- ated by Program. While a user and possibly an LLM have defined this prompt part during application de- velopment, this falls outside the scope of this taxon- omy. Therefore, the Prompt Instruction dimension is not discriminating and categorizes all cases as Pro- gram. It is retained in the taxonomy for completeness and better understandability. Prompt Check describes whether the application em- ploys a review mechanism to control and modify the prompt before invoking the LLM. The same charac- teristics as for the prompt parts are applicable: none: The prompt is used without check. User : The user checks and revises the prompt. LLM : Another LLM component checks or revises the prompt. Program: The application comprises code to check or revise the prompt. E.g., AutoDroid removes personal data, such as names, to ensure privacy before invoking the TaskExecutor; Honeycomb QueryAssistant incorporates a coded mechanism against prompt injection attacks. 13 Table 3: Terms used for prompt parts. Expressions specific to a domain are printed in italics, “examples” indicates a one-shot or few-shot prompt pattern. Terms adopted for the taxonomy are underlined. Source [72] [34] [32] [45] [45] [37] Instruction task description + examples instruction prompt predefined prompt prompt template + examples examples prompt context, i.e., examples [5] [5] [69] [26] education prompt education prompt role and goal + instruction + examples predefined system instruction + domain-specific information State DB schema environment state, scene description dialogue history dialogue history + provided workflow context query results from knowledge graph Task test instance data prompt user prompt user input question SQL query result input task commands user input task prompt (circumscribed) current task the user’s request Most example instances omit prompt checks. There are no examples where a Check is performed by a User or an LLM. 5.2.4. Skills dimensions The Skills dimension captures the types of LLM ca- pabilities that an application utilizes. It is designed as a dimension with six non-mutually exclusive char- acteristics. Skills is decomposed into six specific capabilities: reWrite: The LLM edits or transforms data or text, such as rephrasing, summarizing, reformat- ting, correcting, or replacing values. E.g., My- CrunchGpt SettingsEditor replaces values in JSON files; TruckPlatoon converts measurements into textual explanations. Create: The LLM generates novel output. E.g., LowCode Executing generates substantial bodies of text for tasks like essay writing. conVerse: The application relies on the LLM’s capa- bility to engage in purposeful dialogues with humans. E.g., MyCrunchGpt DesignAssistant asks users for missing parameters; SgpTod PolicyPrompter decides how to react to user inputs and formulates chatbot responses. Inform: The application depends on knowledge that the LLM has acquired during its training, unlike applications that provide all necessary information within the prompt. E.g., MyCrunchGpt Domain- Expert provides expert knowledge on airfoil designs; MatrixProduction relies on built-in knowledge of production processes, such as “a hole is produced by drilling”; LowCode Executing uses its learned knowledge for tasks like essay writing. Reason: The LLM draws conclusions or makes log- ical inferences. E.g., FormulaExplainer in Ex- celCopilot explains the effects of Excel functions in formulas; AutoDroid MemoryGenerators ex- plain the effects of GUI elements in Android apps. Plan: The LLM designs a detailed method or course E.g., Au- of action to achieve a specific goal. toDroid TaskExecutor and WorkplaceRobot TaskPlanning devise action plans to achieve goals. The Plan and Reason characteristics are interrelated, as planning also requires reasoning. The intended handling of these characteristics is to categorize an LLM component as Plan only and understand Plan as implicitly subsuming Reason. The effectiveness of LLMs as components of software applications relies on their commonsense knowledge and their ability to correctly interpret and handle a broad variety of text inputs, including instructions, 14 examples, and code. It is reasonable to assume that a fundamental capability, which might be termed Un- terstand, is leveraged by every LLM component. As it is not distinctive, the taxonomy does not list it explicitly in the Skills dimension. Applying this taxonomy dimension requires users to determine which skills are most relevant and worth highlighting in an LLM component. Given the versa- tility of LLMs, reducing the focus to few predominant skills is necessary to make categorizations distinctive and expressive. 5.2.5. Output-related dimensions Output Format characterizes the format of the LLM’s output. As an output may consist of several parts in diverse formats, this dimension is designed as non- mutually exclusive, same as the Skills dimension. It distinguishes four characteristics that are distinctive and well discernible: FreeText: unstructured natural language text out- put. E.g., TruckPlatoon and MyCrunchGpt DomainExpert generate text output in natural lan- guage; MatrixProduction Manager and Ma- trixProduction Operator produce FreeText ex- planations complementing output in custom formats to be parsed by the application. Item: a single text item from a predefined set of items, such as a class in a classification task. E.g., ProgPrompt ScenarioFeedback outputs either True or False. Code: source code or other highly formalized output that the LLM has learned during its training, such as a programming language, XML, or JSON. E.g., AutoDroid TaskExecutor produces code to steer an Android app; MyCrunchGpt SettingsEditor outputs JSON. Structure: structured, formalized output adhering to a custom format. E.g., LowCode Planning out- puts text in a format that can be displayed as a flow chart; MatrixProduction Manager and Oper- ator produce output in custom formats combined with FreeText explanations. Output Revision indicates whether the application checks or revises the LLM-generated output before utilization. These characteristics and their interpre- tations mirror those in the Prompt Check dimension: none: There is no revision of the LLM output. User : The user revises the LLM output. E.g., the user improves the plan generated by LowCode Planning. LLM : A further LLM component checks or revises the output of the LLM component under considera- tion. Program: Programmed code checks or revises the LLM output. E.g., Honeycomb QueryAssistant corrects the query produced by the LLM before exe- cuting it [7]. There are no instances in the sample set where an- other LLM revises or checks the output of the LLM. Most sample applications do not check or revise the LLM’s output, though several of them parse and transform it. The purpose of the Output Revision dimension is to indicate whether the application in- cludes control or correction mechanisms, rather than just parsing it. Output Consumer addresses the way of utilizing the LLM output: User signifies that the LLM output is presented to a human user. E.g., the text output of TruckPla- toon is intended for humans, as well as the output of MyCrunchGPT DomainExpert. LLM indicates that the output serves as a prompt part in a further LLM invocation. E.g., the knowl- edge base entries generated by an AutoDroid Mem- oryGenerator become part of the prompt for AutoDroid TaskExecutor; the plan output by LowCode Planning serves as a part of the prompt for LowCode Executing. Program describes instances where the LLM output is consumed and processed further by a software com- ponent of the application. E.g., the output of Ma- trixProduction Manager is handled by software systems (including a Manufacturing Execution Sys- tem) which use it to compute prompts for other LLM components. Engine covers scenarios where the LLM output is in- tended for execution on a runtime engine. E.g., the SQL query generated by SgpTod DstPrompter is 15 processed by a SQL interpreter; a part of the output of MatrixProduction Operator is executed by automation modules. Although applications may parse and transform the LLM output before use, the Output Consumer di- mension is meant to identify the ultimate consumer, such as an execution engine, rather than an interme- diary parser or transformation code. When applica- tions divide the LLM output into parts for different consumers, users applying the taxonomy need to de- termine which consumer is most relevant, since this dimension is designed to be mutually exclusive. 5.3. Evaluation Figure 2 displays the number of occurrences of char- It must acteristics within the example instances. be noted, however, that these do not reflect actual frequencies, as similar LLM components within the same application are aggregated together, indicated by symbols ∗ and 2 in figure 1. Furthermore, Ex- celCopilot likely includes occurrences of Prompt Check and Output Revision which are not counted due to insufficient system documentation. We evaluate the taxonomy against commonly ac- cepted quality criteria: comprehensiveness, robust- ness, conciseness, mutual exclusiveness, explanatory power, and extensibility [58, 42]. The taxonomy encompasses all example instances including those that were not considered during its development. This demonstrates comprehensiveness. As figure 1 shows, all example instances have unique categoriza- tions, supporting the taxonomy’s robustness. This not only indicates that the dimensions and charac- teristics are distinctive for the domain, but also high- lights the wide variety possible in this field. Concise- ness demands that the taxonomy uses the minimum number of dimensions and characteristics. The tax- onomy gains conciseness by identifying relatively few and abstract characteristics within each dimension. However, it does not adhere to the related subcri- terion that each characteristic must be present in at least one investigated instance [54]. Unoccupied char- acteristics are retained for dimensions whose char- acteristics were derived conceptually, specifically, for the Prompt dimensions, the Output Revision dimen- sion, and the Data Function dimension, enhancing the taxonomy’s ability to illustrate design options and inspire novel uses for LLM integrations in ap- plications. Some dimensions are constructed in par- allel, sharing common sets of characteristics. While this affects conciseness, it makes the taxonomy easier to understand and apply. As is often seen in tax- onomy development [54], we deliberately waived the requirement for mutual exclusiveness for some di- mensions, specifically the Output Format and Skills dimensions. In the context of this taxonomy, these can equivalently be understood as a set of of six and four binary dimensions respectively, each divided into characteristics “yes” and “no”. However, framing them as a single dimension with non-mutually exclu- sive characteristics seems more intuitive. Metadimensions structure the taxonomy, and most of the characteristics are illustrated through exam- ples. These measures are recognized for enhancing the explanatory power of a taxonomy [58]. The taxonomy’s flat structure allows for the easy addition of dimensions and characteristics, indicating that its extensibility is good. Potential extensions and fur- ther aspects of the taxonomy, including its usefulness and ease of use, are discussed in section 6. We visualize the taxonomy (or, strictly speaking, cat- egorized instances) in a compact form using feature vectors with characteristics abbreviated to single- letter codes. This approach has a drawback, as it requires referencing a legend. Additionally, non- applicable characteristics in mutually exclusive di- mensions are not visible, which means the design space is not completely shown. However, the com- pactness of the representation allows LLM compo- nents within a common application to be grouped closely, so that an LLM-integrated application can be perceived as a unit without appearing convoluted. This is a significant advantage for our purposes. 6. Discussion The discussion first focuses on the taxonomy’s appli- cability and ease of use before considering its overall usefulness. 16 Invocation (cid:122) (cid:123) (cid:125)(cid:124) Inter. Freq. Logic UI Function (cid:125)(cid:124) (cid:122) (cid:123) Data (cid:122) Instr. Prompt (cid:125)(cid:124) State Task (cid:123) Check Skills (cid:125)(cid:124) (cid:122) (cid:123) Output Format (cid:122) (cid:122) (cid:125)(cid:124) (cid:123) Revision Consumer Output (cid:125)(cid:124) (cid:123) A C D I S C A I O B R W B U L P U L P U L P U L P W C V I R P F I C S U L P U L P E 8 9 4 5 16 8 13 5 2 2 2 0 0 0 0 21 0 2 17 11 3 7 0 0 2 3 1 4 4 7 8 10 4 6 8 1 0 1 5 3 3 10 Figure 2: Occurrences of characteristics in the sample set of LLM-integrated applications. 6.1. Applicability and ease of use The taxonomy was effectively applied to LLM- integrated applications based on research papers, source code blog posts, recorded software demonstra- tions, and developer experiences. The analysis of LowCode revealed it to be a prompt definition tool combined with an LLM-based chatbot, which devi- ates from the strict definition of an LLM-integrated application. Still, the taxonomy provided an effective categorization and led to a clear understanding of the system’s architecture. Obviously, the ease of categorization depends on the clarity and comprehensiveness of the available infor- mation, which varies across analyzed systems. An- alyzing applications of LLMs in novel and uncom- mon domains can be challenging. While these papers present inspiring and innovative ideas for LLM inte- gration, such as MyCrunchGpt and TruckPla- toon, they may prioritize explaining the application area and struggle to detail the technical aspects of the LLM integration. A taxonomy for LLM-integrated applications can guide and facilitate the writing pro- cess and lead to more standardized and comparable descriptions. Applying the taxonomy is often more straightforward for research-focused systems. Omitting the com- plexities required for real-world applications, such as prompt checks and output revisions, their architec- tures are simpler and easier to describe. A taxonomy can point out such omissions. A fundamental challenge in applying the taxonomy arises from the inherent versatility of LLMs, which allows to define LLM components serving multiple purposes. This is exemplified by SgpTod Poli- cyPrompter, where the prompt is designed to pro- duce a structure with two distinct outcomes (a class label and a chatbot response), and similarly by Ma- trixProduction, as detailed section 4.2. Draw- ing an analogy to “function overloading” in classical programming, such LLM components can be termed “overloaded LLM components”. A taxonomy can handle overloaded LLM components in several ways: (1) define more dimensions as non- mutually exclusive, (2) label overloaded LLM compo- nents as “overloaded” without a more detailed catego- rization, or (3) categorize them by their predominant purpose or output. While the first approach allows for the most precise categorization, it complicates the taxonomy. Moreover, it will likely result in nearly all characteristics being marked for some LLM compo- nents, which is ultimately not helpful. The second approach simplifies categorization but sacrifices much detail. Our taxonomy adopts the third approach, en- forcing simplification and abstraction in descriptions of overloaded LLM components while retaining es- sential detail. The taxonomy can easily be extended to include approach (2) as an additional binary di- mension. 6.2. Usefulness The search for instances of LLM-integrated appli- cations uncovered activities across various domains. Substantial research involving LLM integrations, of- ten driven by theoretical interests, is notable in robot task planning [37, 51, 61, 33, 63] and in the TOD field [23, 71, 4, 6, 56]. Research exploring LLM po- tentials from a more practical perspective can be found in novel domains, such as industrial produc- tion [69, 26] and other technical areas [28, 70]. Fur- 17 thermore, developers of commercial LLM-based ap- plications are beginning to communicate their efforts and challenges [44, 7]. The taxonomy has been ap- plied to example instances from these and additional areas. This demonstrates its potential as a common, unified framework for describing LLM-integrated ap- plications, facilitating the comparison and sharing of development knowledge between researchers and practitioners across various domains. When applying the taxonomy to the example in- stances, it proved to be effective and useful as an analytical lens. Descriptions of LLM-integrated ap- plications commonly explain background information and details of the application domain in addition to its LLM integration. When used as an analytical lens, the taxonomy quickly directs the analysis to- wards the aspects of LLM integration, abstracting from the specificities of the domain. The taxonomy describes how LLM capabilities can be leveraged in software systems, offers inspiration for LLM-based functions, and outlines options for their implementation as follows. The Skills dimension out- lines the range of capabilities an LLM can contribute to an application through a concise set of characteris- tics, while the Function dimension suggests potential uses, further supported by the Interaction dimension. The Output Type dimension indicates options for en- coding the output of an LLM in formats beyond plain text, making it processable by software. The Output Consumer dimension illustrates the diverse ways to utilize or act upon LLM output. Thus, the taxonomy, as intended, spans a design space for LLM integra- tions. The sampled LLM-integrated applications showcase the creativity of researchers and developers in ap- plying and exploiting the potentials of LLMs, rang- ing from straightforward solutions (e.g., TruckPla- toon) to highly sophisticated and technically com- plex ones (e.g., AutoDroid). When using the tax- onomy to inspire innovative uses of LLMs, we recom- mend supplementing it with descriptions of example applications to enhance its illustrativeness. The char- acteristics of the Skills dimension are derived prag- matically from the investigated example instances. While they do not claim to be exhaustive or deeply 18 rooted in LLM theory or cognitive science, they add relevant details to the categorizations and illustrate design options and potentials for using LLMs as soft- ware components. It emerged as a key insight of this research that, rather than analyzing an LLM-integrated application in whole, analysis should start with the identifica- tion and description of its distinct LLM components. This is essential for gaining a clear understanding of how the application utilizes the capabilities of LLMs. The LLM-integrated application then manifests as a combination of its LLM components. As shown in fig- ure 1, the visualization effectively displays both the quantity and the variety of LLM components in an LLM-integrated application. LLM components interact through prompt chaining, where one LLM component’s output feeds into an- other’s input [67]. When an LLM-integrated applica- tion involves such an interaction, the taxonomy rep- resents it as an LLM characteristic within a Prompt dimension. The taxonomy can capture the variance in these interactions. For instance, in AutoDroid TaskExecutor and LowCode Executing, the LLM characteristic appears in the Prompt State di- mension, because their prompt components (knowl- edge base excerpts and prompt definition, respec- tively) are generated by other LLM components in a preparatory stage. In contrast, the LLM character- istic appears in the Prompt Task dimension for Ma- trixProduction Operator, because its prompt part is generated individually by the MatrixPro- duction Manager almost immediately before use. that cover Taxonomy dimensions entire LLM- integrated applications may be useful. Given their complexity, these dimensions should be designed based on a broader range of examples, which will only become available as more LLM-integrated applica- tions are developed and their architectures disclosed in the future. Extensions to the taxonomy could also include dimensions for describing the structure of prompts in more detail, as well as dimensions ad- dressing characteristics of the language models used. Table 4: LLM usage in the sample instances. “Evals” indicates evaluations of various LLMs. Used or best LLM Evals Comments GPT-3.5 GPT-3.5-turbo GPT-3.5 yes GPT-4 far too slow then awaiting the publication of GPT-4 Application Honeycomb LowCode MyCrunchGpt MatrixProduction text-davinci-003 WorkplaceRobot AutoDroid ProgPrompt FactoryAssistants GPT-3.5 GPT-3.5 SgpTod GPT-3.5-turbo TruckPlatoon N/A ExcelCopilot GPT-3 GPT-4 GPT-3 yes GPT-4 best for tasks requiring many steps CODEX better, but access limits prohibitive yes GPT-3.5 best more often than others combined combined LLMs in Copilot for Microsoft 365 [43] 7. Conclusion This paper investigates the use of LLMs as soft- ware components. Its perspective differs from cur- rent software engineering research, which investigates LLMs as tools for software development [14, 22] and from research examining LLMs as autonomous agents [11, 62, 57, 21]. This paper defines the concept of an LLM component as a software component that re- alizes its functionality by invoking an LLM. While LLM components implicitly appear in various works, termed, for example, “prompters”, “prompted LLM”, “prompt module”, or “module” [30, 71, 6, 7], to our knowledge, this concept has not yet been formalized or systematically investigated. The main contribution of this study is a taxonomy for the analysis and description of LLM components, extending to LLM-integrated applications by charac- terizing them as combinations of LLM components. In addition to the dimensions and characteristics of the taxonomy, the study contributes a taxonomy vi- sualization based on feature vectors, which is more compact than the established visualizations such as morphological boxes [55] or radar charts. It repre- sents an LLM-integrated application as one visual en- tity in a tabular format, with its LLM components displayed as rows. The taxonomy was constructed using established methods, based on a set of example instances, and evaluated with a new set of example instances. The combined samples exhibit broad variation along the identified dimensions. For some instances, informa- tion was not available, necessitating speculative in- terpretation. However, since the sample is used for identifying options rather than quantitative analysis, this issue and the representativeness of the sample are not primary concerns. The evaluation was con- ducted by the developer of the taxonomy, consistent with recent related work [21, 52, 48]. Using a new sample for evaluation strengthens the validity of the results. A further significant contribution of the paper is a systematic overview of a sample of LLM-integrated applications across various industrial and technical domains, illustrating a spectrum of conceptual ideas and implementation options. As the examples show, LLM components can re- place traditionally coded functions in software sys- tems and enable novel use cases. However, practi- cal challenges persist. Developers report that new software engineering methods are required, e.g., for managing prompts as software assets and for test- ing and monitoring applications. For instance, the costs of LLM invocations prohibit the extensive au- tomated testing that is standard in software devel- opment practice [44, 7]. Challenges also arise from the inherent indeterminism and uncontrollability of LLMs. Small variations in prompts can lead to differ- ences in outputs, while automated output processing 19 in LLM-integrated applications requires the output to adhere to a specified format. Furthermore, the deployment mode of LLMs, whether local (on the same hardware as the ap- plication) or remote, managed privately or offered as Language-Models-as-a-Service (LMaaS), has im- pact on performance and usability. Table 4 gives an overview of the LLMs used in our sample of appli- cations. Where papers report evaluations of mul- tiple LLMs, the table displays the chosen or best- performing LLM. Although not representative, the table provides some insights. LMaaS dominates, likely due to its convenience, but more importantly, due to the superior performance of the provided LLMs. Concerns regarding LMaaS include privacy, as sensi- tive data might be transmitted to the LLM through the prompt [64], and service quality, i.e., reliability, availability, and costs. Costs typically depend on the quantity of processed tokens. This quantity also af- fects latency, which denotes the processing time of an LLM invocation. A further important factor for latency is the size of the LLM, with larger models being slower [7]. When building LLM-based applications for real- world use, the reliability and availability of an LMaaS are crucial. Availability depends not only on the technical stability of the service, but also on factors such as increased latency during high usage periods or usage restrictions imposed by the provider of an LMaaS, as reported for ProgPrompt [51]. Beyond technical aspects, the reliability of an LMaaS also en- compasses its behavior. For instance, providers might modify a model to enhance its security, potentially impacting applications that rely on it. Despite practical challenges, integrating LLMs into systems has the potential to alter the way software is constructed and the types of systems that can be realized. Prompts are central to the functioning of LLM components which pose specific requirements such as strict format adherence. Therefore, an im- portant direction for future research will be prompt engineering specifically tailored for LLM-integrated applications. In future work, the taxonomy will be extended to distinguish finer-grained parts of prompts, allowing a more detailed description and comparison of prompts and related experimental results. Initial studies share results on the format-following behavior of LLMs [68] as a subtopic of instruction-following [73], derived with synthetic benchmark data. It is necessary to complement their results with experiments using data and tasks from real application development projects because, in the early stages of this field, synthetic benchmarks may fail to cover relevant aspects within the wide range of possible options. Another crucial research direction involves exploring how LLM char- acteristics correspond to specific tasks, such as de- termining the optimal LLM size for intent detection tasks. The taxonomy developed in this study can sys- tematize such experiments and their outcomes. Ad- ditionally, it provides a structured framework for de- lineating design choices in LLM components, making it a valuable addition to future training materials. Acknowledgements Special thanks to Antonia Weber and Constantin We- ber for proofreading and providing insightful and con- structive comments. References [1] Eleni Adamopoulou and Lefteris Moussiades. An Overview of Chatbot Technology. In Ilias Ma- glogiannis, Lazaros Iliadis, and Elias Pimeni- dis, editors, Artificial Intelligence Applications and Innovations, IFIP Advances in Information and Communication Technology, pages 373–383, Cham, 2020. Springer International Publishing. doi:10.1007/978-3-030-49186-4_31. [2] Sebastian Bader, Erich Barnstedt, Heinz Be- denbender, Bernd Berres, Meik Billmann, and Marko Ristin. Details of the asset adminis- tration shell-part 1: The exchange of informa- tion between partners in the value chain of in- dustrie 4.0 (version 3.0 rc02). Working Paper, Berlin: Federal Ministry for Economic Affairs 20 and Climate Action (BMWK), 2022. doi.org/ 10.21256/zhaw-27075. Soft Computing, 151:111165, January 2024. doi:10.1016/j.asoc.2023.111165. [3] Marcos Baez, Florian Daniel, Fabio Casati, and Boualem Benatallah. Chatbot integration in few patterns. IEEE Internet Computing, pages 1–1, 2020. doi:10.1109/MIC.2020.3024605. [4] Tom Bocklisch, Thomas Werkmeister, Task- Daksh Varshneya, and Alan Nichol. Oriented Dialogue with In-Context Learn- ing. (arXiv:2402.12234), February 2024. doi:10.48550/arXiv.2402.12234. [5] Yuzhe Cai, Shaoguang Mao, Wenshan Wu, Ze- hua Wang, Yaobo Liang, Tao Ge, Chenfei Wu, Wang You, Ting Song, Yan Xia, Jonathan Tien, and Nan Duan. Low-code LLM: Visual Pro- gramming over LLMs. (arXiv:2304.08103), April 2023. doi:10.48550/arXiv.2304.08103. [6] Lang Cao. DiagGPT: An LLM-based Chatbot with Automatic Topic Management for Task- Oriented Dialogue. (arXiv:2308.08043), August 2023. doi:10.48550/arXiv.2308.08043. [7] Phillip Carter. All the Hard Stuff No- body Talks About When Building Prod- ucts with LLMs. Honeycomb, May 2023. https://www.honeycomb.io/blog/ hard-stuff-nobody-talks-about-llm. [8] Phillip Carter. So We Shipped an AI Prod- Honeycomb, Octo- uct. Did It Work? ber 2023. https://www.honeycomb.io/blog/ we-shipped-ai-product. [9] Banghao Chen, Zhaofeng Zhang, Nicolas Langrené, Unleash- and Shengxin Zhu. ing the potential of prompt engineering in Large Language Models: A comprehensive review. (arXiv:2310.14735), October 2023. doi:10.48550/arXiv.2310.14735. [10] Wang Chen, Yan-yi Liu, Tie-zheng Guo, Da- peng Li, Tao He, Li Zhi, Qing-wen Yang, Hui-han Wang, and Ying-you Wen. Sys- industry appli- tems engineering issues cations of Applied large language model. for 21 [11] Yuheng Cheng, Ceyao Zhang, Zhengwen Zhang, Xiangrui Meng, Sirui Hong, Wenhao Li, Zihao Wang, Zekai Wang, Feng Yin, Junhua Zhao, and Xiuqiang He. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. (arXiv:2401.03428), January 2024. doi:10.48550/arXiv.2401.03428. [12] Silvia Colabianchi, Andrea Tedeschi, and Francesco Costantino. Human-technology in- tegration with industrial conversational agents: A conceptual architecture and a taxonomy for manufacturing. Journal of Industrial Infor- mation Integration, 35:100510, October 2023. doi:10.1016/j.jii.2023.100510. [13] Jonathan Evertz, Merlin Chlosta, Lea Schön- herr, and Thorsten Eisenhofer. Whispers in the Machine: Confidentiality in LLM-integrated Systems. (arXiv:2402.06922), February 2024. doi:10.48550/arXiv.2402.06922. [14] Angela Fan, Beliz Gokkaya, Mark Harman, Mitya Lyubarskiy, Shubho Sengupta, Shin Yoo, and Jie M. Zhang. Large Language Models for Software Engineering: Survey and Open Problems. (arXiv:2310.03533), November 2023. doi:10.48550/arXiv.2310.03533. [15] Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. Recommender Systems in the Era of Large Language Models (LLMs). (arXiv:2307.02046), August 2023. doi:10.48550/arXiv.2307.02046. [16] David Fortin. Microsoft Copilot in Excel: What It Can and Can’t Do. YouTube, Jan- uary 2024. https://www.youtube.com/watch? v=-fsu9IXMZvo. [17] Martin Fowler. Patterns of Enterprise Applica- tion Architecture. 2002. ISBN 978-0-321-12742- 6. [18] Shirley Gregor. The nature of theory in infor- mation systems. MIS quarterly, pages 611–642, 2006. doi:10.2307/25148742. [19] Yanchu Guan, Dong Wang, Zhixuan Chu, Shiyu Wang, Feiyue Ni, Ruihua Song, Longfei Li, Jin- jie Gu, and Chenyi Zhuang. Intelligent Vir- tual Assistants with LLM-based Process Au- tomation. (arXiv:2312.06677), December 2023. doi:10.48550/arXiv.2312.06677. [20] Muhammad Usman Hadi, Qasem Al Tashi, Rizwan Qureshi, Abbas Shah, Amgad Muneer, Muhammad Irfan, Anas Zafar, Muhammad Bi- lal Shaikh, Naveed Akhtar, Jia Wu, and Seyedali Mirjalili. Large Language Models: A Compre- hensive Survey of its Applications, Challenges, Limitations, and Future Prospects, September 2023. doi:10.36227/techrxiv.23589741.v3. [21] Thorsten Händler. A Taxonomy for Au- tonomous LLM-Powered Multi-Agent Architec- tures:. In Proceedings of the 15th Interna- tional Joint Conference on Knowledge Discov- ery, Knowledge Engineering and Knowledge Management, pages 85–98, Rome, Italy, 2023. SCITEPRESS - Science and Technology Publi- cations. doi:10.5220/0012239100003598. [22] Xinyi Hou, Yanjie Zhao, Yue Liu, Zhou Yang, Kailong Wang, Li Li, Xiapu Luo, David Lo, John Grundy, and Haoyu Wang. Large Language Models for Software Engineering: A Systematic Literature Review. (arXiv:2308.10620), Septem- ber 2023. doi:10.48550/arXiv.2308.10620. [23] Vojtěch Hudeček and Ondrej Dusek. Are Large Language Models All You Need for Task- In Svetlana Stoyanchev, Oriented Dialogue? Shafiq Joty, David Schlangen, Ondrej Dusek, Casey Kennington, and Malihe Alikhani, edi- tors, Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Di- alogue, pages 216–228, Prague, Czechia, Septem- ber 2023. Association for Computational Lin- guistics. doi:10.18653/v1/2023.sigdial-1.21. [24] Kevin Maik Jablonka, Qianxiang Ai, Alexander Al-Feghali, Shruti Badhwar, Joshua D. Bocarsly, Andres M. Bran, Stefan Bringuier, Catherine L. Brinson, Kamal Choudhary, Defne Circi, Sam Cox, Wibe A. de Jong, Matthew L. Evans, Nico- las Gastellu, Jerome Genzling, María Victoria Gil, Ankur K. Gupta, Zhi Hong, Alishba Im- ran, Sabine Kruschwitz, Anne Labarre, Jakub Lála, Tao Liu, Steven Ma, Sauradeep Majum- dar, Garrett W. Merz, Nicolas Moitessier, Elias Moubarak, Beatriz Mouriño, Brenden Pelkie, Michael Pieler, Mayk Caldas Ramos, Bojana Ranković, Samuel Rodriques, Jacob Sanders, Philippe Schwaller, Marcus Schwarting, Jiale Shi, Berend Smit, Ben Smith, Joren Van Herck, Christoph Völker, Logan Ward, Sean War- ren, Benjamin Weiser, Sylvester Zhang, Xiaoqi Zhang, Ghezal Ahmad Zia, Aristana Scour- tas, K. Schmidt, Ian Foster, Andrew White, and Ben Blaiszik. 14 examples of how LLMs can transform materials science and chem- istry: A reflection on a large language model hackathon. Digital Discovery, 2(5):1233–1250, 2023. doi:10.1039/D3DD00113J. [25] Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. Challenges and Applica- tions of Large Language Models, July 2023. doi:10.48550/arXiv.2307.10169. [26] Samuel Kernan Freire, Mina Foosherian, Chao- fan Wang, and Evangelos Niforatos. Harnessing Large Language Models for Cognitive Assistants in Factories. In Proceedings of the 5th Interna- tional Conference on Conversational User Inter- faces, CUI ’23, pages 1–6, New York, NY, USA, July 2023. Association for Computing Machin- ery. doi:10.1145/3571884.3604313. [27] Anis Koubaa, Wadii Boulila, Lahouari Ghouti, Ayyub Alzahem, and Shahid Latif. Explor- ing ChatGPT Capabilities and Limitations: A Survey. IEEE Access, 11:118698–118721, 2023. doi:10.1109/ACCESS.2023.3326474. [28] Varun Kumar, Leonard Gleyzer, Adar Ka- hana, Khemraj Shukla, and George Em Karni- 22 adakis. MyCrunchGPT: A LLM Assisted Frame- work for Scientific Machine Learning. Jour- nal of Machine Learning for Modeling and Computing, 4(4), 2023. doi.org/10.1615/ JMachLearnModelComput.2023049518. [29] Dennis Jan Kundisch, Muntermann, Anna Maria Oberländer, Daniel Rau, Maxi- milian Röglinger, Thorsten Schoormann, and Daniel Szopinski. An Update for Taxonomy Designers. Business & Information Systems Engineering, 2022. doi:10.1007/s12599-021-00723-x. 64(4):421–439, August Prompted LLMs as Jongho [30] Gibbeum Lee, Volker Hartmann, and Kang- Park, Dimitris Papailiopoulos, wook Lee. chatbot modules for long open-domain conversation. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Findings of the as- sociation for computational linguistics: ACL 2023, pages 4536–4554, Toronto, Canada, July 2023. Association for Computational Linguistics. doi:10.18653/v1/2023.findings-acl.277. [31] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zheng- bao Jiang, Hiroaki Hayashi, and Graham Neu- big. Pre-train, Prompt, and Predict: A Sys- tematic Survey of Prompting Methods in Nat- ural Language Processing. ACM Comput- ing Surveys, 55(9):195:1–195:35, January 2023. doi:10.1145/3560815. [32] Yi Liu, Gelei Deng, Yuekang Li, Kailong Wang, Tianwei Zhang, Yepang Liu, Haoyu Wang, Yan Zheng, and Yang Liu. Prompt Injection at- tack against LLM-integrated Applications, June 2023. doi:10.48550/arXiv.2306.05499. [33] Yuchen Liu, Luigi Palmieri, Sebastian Ilche Georgievski, and Marco Aiello. Koch, DELTA: Decomposed Efficient Long-Term Robot Task Planning using Large Language Models. (arXiv:2404.03275), April 2024. doi:10.48550/arXiv.2404.03275. [34] Yupei Liu, Yuqi Jia, Runpeng Geng, Jinyuan Jia, and Neil Zhenqiang Gong. Prompt Injec- tion Attacks and Defenses in LLM-Integrated 23 Applications. (arXiv:2310.12815), October 2023. doi:10.48550/arXiv.2310.12815. [35] Shaoguang Mao, Qiufeng Yin, Yuzhe Cai, https: and Dan Qiao. //github.com/chenfei-wu/TaskMatrix/ tree/main/LowCodeLLM, May 2023. LowCodeLLM. [36] Scott McLean, Gemma J. M. Read, Jason Thompson, Chris Baber, Neville A. Stanton, and Paul M. Salmon. The risks associated with Ar- tificial General Intelligence: A systematic re- view. Journal of Experimental & Theoretical Artificial Intelligence, 35(5):649–663, July 2023. doi:10.1080/0952813X.2021.1964003. [37] Oier Mees, Jessica Borja-Diaz, and Wolfram Burgard. Grounding Language with Visual Af- In 2023 fordances over Unstructured Data. IEEE International Conference on Robotics and Automation (ICRA), pages 11576–11582, London, United Kingdom, May 2023. IEEE. doi:10.1109/ICRA48891.2023.10160396. [38] Grégoire Mialon, Roberto Dessì, Maria Lomeli, Christoforos Nalmpantis, Ram Pa- sunuru, Roberta Raileanu, Baptiste Rozière, Timo Schick, Jane Dwivedi-Yu, Asli Ce- likyilmaz, Edouard Grave, Yann LeCun, and Thomas Scialom. Augmented Lan- guage Models: A Survey, February 2023. doi:10.48550/arXiv.2302.07842. [39] Melanie Mitchell. ture of artificial general ence, doi:10.1126/science.ado7069. intelligence. 383(6689):eado7069, March Debates on the na- Sci- 2024. [40] Quim Motger, Xavier Franch, and Jordi Marco. Survey, Software-Based Dialogue Systems: Taxonomy, and Challenges. ACM Comput- ing Surveys, 55(5):91:1–91:42, December 2022. doi:10.1145/3527450. [41] Fiona Fui-Hoon Nah, Ruilin Zheng, Jingyuan Cai, Keng Siau, and Langtao Chen. Gen- erative AI and ChatGPT: Applications, chal- lenges, and AI-human collaboration. Jour- nal of Information Technology Case and Ap- plication Research, 25(3):277–304, July 2023. doi:10.1080/15228053.2023.2233814. [42] Robert C Nickerson, Upkar Varshney, and taxon- Jan Muntermann. omy development and its application in in- formation systems. European Journal of In- formation Systems, 22(3):336–359, May 2013. doi:10.1057/ejis.2012.26. A method for [43] Camille Pack, Cern McAtee, Samantha Robert- son, Dan Brown, Aditi Srivastava, and Kweku Ako-Adjei. Microsoft Copilot for Microsoft 365 overview. https://learn.microsoft. com/en-us/copilot/microsoft-365/ microsoft-365-copilot-overview, 2024. March Sumit Gulwani, [44] Chris Parnin, Gustavo Soares, Rahul Pan- dita, and Austin Z. Henley. Building Your Own Prod- uct Copilot: Challenges, Opportunities, and Needs. (arXiv:2312.14231), December 2023. doi:10.48550/arXiv.2312.14231. Jessica Rich, [45] Rodrigo Pedro, Daniel Castro, Paulo Car- From Prompt In- reira, and Nuno Santos. jections to SQL Injection Attacks: How Pro- tected is Your LLM-Integrated Web Appli- cation? (arXiv:2308.01990), August 2023. doi:10.48550/arXiv.2308.01990. [46] Ken Peffers, Tuure Tuunanen, Marcus A. Rothenberger, and Samir Chatterjee. A De- sign Science Research Methodology for Infor- mation Systems Research. Journal of Man- agement Information Systems, 24(3):45–77, De- cember 2007. ISSN 0742-1222, 1557-928X. doi:10.2753/MIS0742-1222240302. [47] Mohaimenul Azam Khan Raiaan, Md. Sad- dam Hossain Mukta, Kaniz Fatema, Nur Mo- hammad Fahad, Sadman Sakib, Most Mar- Jubaer Ahmad, Mo- ufatul Jannat Mim, hammed Eunus Ali, and Sami Azam. A Review on Large Language Models: Architectures, Ap- plications, Taxonomies, Open Issues and Chal- 24 lenges. doi:10.1109/ACCESS.2024.3365742. IEEE Access, 12:26839–26874, 2024. [48] Jack Daniel Rittelmeyer and Kurt Sandkuhl. Morphological Box for AI Solutions: Evalua- tion and Refinement with a Taxonomy Develop- ment Method. In Knut Hinkelmann, Francisco J. López-Pellicer, and Andrea Polini, editors, Per- spectives in Business Informatics Research, Lec- ture Notes in Business Information Process- ing, pages 145–157, Cham, 2023. Springer Na- ture Switzerland. doi:10.1007/978-3-031-43126- 5_11. [49] Shubhra Kanti Karmaker Santu and Dongji TELeR: A General Taxonomy of for Benchmarking Complex (arXiv:2305.11430), October 2023. Feng. LLM Prompts Tasks. doi:10.48550/arXiv.2305.11430. [50] Thorsten Schoormann, Frederik Möller, and Daniel Szopinski. Exploring Purposes of Us- In Proceedings of the Inter- ing Taxonomies. national Conference on Wirtschaftsinformatik (WI), Nuernberg, Germany, February 2022. [51] Ishika Singh, Valts Blukis, Arsalan Mousa- vian, Ankit Goyal, Danfei Xu, Jonathan Trem- blay, Dieter Fox, Jesse Thomason, and Ani- mesh Garg. ProgPrompt: Generating Situated Robot Task Plans using Large Language Mod- els. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523– 11530, London, United Kingdom, May 2023. IEEE. doi:10.1109/ICRA48891.2023.10161317. [52] Gero Strobel, Leonardo Banh, Frederik Möller, and Thorsten Schoormann. Exploring Gener- ative Artificial Intelligence: A Taxonomy and Types. In Proceedings of the 57th Hawaii Inter- national Conference on System Sciences, Hon- olulu, Hawaii, January 2024. https://hdl. handle.net/10125/106930. [53] Hendrik Strobelt, Albert Webson, Victor Sanh, Benjamin Hoover, Johanna Beyer, Hanspeter Pfister, and Alexander M. Rush. Interac- tive and Visual Prompt Engineering for Ad- hoc Task Adaptation With Large Language Models. IEEE Transactions on Visualization and Computer Graphics, pages 1–11, 2022. doi:10.1109/TVCG.2022.3209479. Effective Invocation Methods of Massive LLM Services. (arXiv:2402.03408), February 2024. doi:10.48550/arXiv.2402.03408. [54] Daniel Szopinski, Thorsten Schoormann, and Dennis Kundisch. Criteria as a Prelude for Guid- ing Taxonomy Evaluation. In Proceedings of the 53rd Hawaii International Conference on Sys- tem Sciences, 2020. https://hdl.handle.net/ 10125/64364. [55] Daniel Szopinski, Thorsten Schoormann, and Visualize different: To- Dennis Kundisch. researching the fit between taxon- wards omy visualizations and taxonomy tasks. In Tagungsband Der 15. Internationalen Tagung Wirtschaftsinformatik (WI 2020), Potsdam, 2020. doi:10.30844/wi_2020_k9-szopinski. [56] Manisha Thakkar and Nitin Pise. Unified Ap- proach for Scalable Task-Oriented Dialogue Sys- tem. International Journal of Advanced Com- puter Science and Applications, 15(4), 2024. doi:10.14569/IJACSA.2024.01504108. [57] Oguzhan Topsakal and Tahir Cetin Akinci. Cre- ating Large Language Model Applications Uti- lizing Langchain: A Primer on Developing LLM Apps Fast. In International Conference on Applied Engineering and Natural Sciences, vol- ume 1, pages 1050–1056, 2023. [58] Michael Unterkalmsteiner and Waleed Adbeen. A compendium and evaluation of taxonomy quality attributes. Expert Systems, 40(1): e13098, 2023. doi:10.1111/exsy.13098. [59] Bryan Wang, Gang Li, and Yang Li. En- Interaction with Mo- abling Conversational In bile UI using Large Language Models. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, pages 1–17, New York, NY, USA, April 2023. Association for Computing Machinery. doi:10.1145/3544548.3580895. [61] Jun Wang, Guocheng He, and Yiannis Kan- Safe Task Planning for Language- taros. Instructed Multi-Robot Systems using Confor- mal Prediction. (arXiv:2402.15368), February 2024. doi:10.48550/arXiv.2402.15368. [62] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, and Jirong Wen. A survey on large language model based autonomous agents. Frontiers of Com- puter Science, 18(6):186345, March 2024. doi:10.1007/s11704-024-40231-1. [63] Shu Wang, Muzhi Han, Ziyuan Jiao, Zeyu Zhang, Ying Nian Wu, Song-Chun Zhu, and Hangxin Liu. LLM3:Large Language Model- based Task and Motion Planning with Motion Failure Reasoning. (arXiv:2403.11552), March 2024. doi:10.48550/arXiv.2403.11552. [64] Hao Wen, Yuanchun Li, Guohong Liu, Shan- hui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. Em- powering LLM to use Smartphone for Intelligent Task Automation. (arXiv:2308.15272), Septem- ber 2023. doi:10.48550/arXiv.2308.15272. [65] Hao Wen, Yuanchun Li, and Sean KiteFly- Kid. MobileLLM/AutoDroid. Mobile LLM, Jan- uary 2024. https://github.com/MobileLLM/ AutoDroid. [66] Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, and Dou- Jesse Spencer-Smith, glas C. Schmidt. A Prompt Pattern Cat- alog to Enhance Prompt Engineering with ChatGPT. (arXiv:2302.11382), February 2023. doi:10.48550/arXiv.2302.11382. [60] Can Wang, Bolin Zhang, Dianbo Sui, Zhiying Tu, Xiaoyu Liu, and Jiabao Kang. A Survey on [67] Tongshuang Wu, Michael Terry, and Car- rie Jun Cai. AI Chains: Transparent and 25 Instruction- and Le Hou. Denny Zhou, Following Evaluation for Large Language Mod- els. (arXiv:2311.07911), November 2023. doi:10.48550/arXiv.2311.07911. Controllable Human-AI Interaction by Chain- ing Large Language Model Prompts. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, CHI ’22, pages 1–22, New York, NY, USA, April 2022. Association for Computing Machinery. doi:10.1145/3491102.3517582. [68] Congying Xia, Chen Xing, Jiangshu Du, Xinyi Yang, Yihao Feng, Ran Xu, Wenpeng Yin, and Caiming Xiong. FOFO: A Benchmark to Evaluate LLMs’ Format-Following Capa- bility. (arXiv:2402.18667), February 2024. doi:10.48550/arXiv.2402.18667. [69] Yuchen Xia, Manthan Shenoy, Nasser Jazdi, and Michael Weyrich. Towards autonomous system: Flexible modular production sys- language model tem enhanced with large agents. In 2023 IEEE 28th International Con- ference on Emerging Technologies and Fac- tory Automation (ETFA), pages 1–8, 2023. doi:10.1109/ETFA54631.2023.10275362. [70] I. de Zarzà, J. de Curtò, Gemma Roig, and Carlos T. Calafate. LLM Adaptive PID Control for B5G Truck Platooning Sys- tems. Sensors, 23(13):5899, January 2023. doi:10.3390/s23135899. [71] Xiaoying Zhang, Baolin Peng, Kun Li, Jingyan SGP-TOD: Build- Zhou, and Helen Meng. ing Task Bots Effortlessly via Schema-Guided LLM Prompting. (arXiv:2305.09067), May 2023. doi:10.48550/arXiv.2305.09067. [72] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A Survey of Large Lan- guage Models. (arXiv:2303.18223), May 2023. doi:10.48550/arXiv.2303.18223. [73] Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, 26 

synthetic_cpt

Role_of_Data_Augmentation_Strategies_in_Knowledge_Distillation_for_Wearable_Sensor_Data.pdf

IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 1 Role of Data Augmentation Strategies in Knowledge Distillation for Wearable Sensor Data Eun Som Jeon, Student Member, IEEE, Anirudh Som, Ankita Shukla, Kristina Hasanaj, Matthew P. Buman, and Pavan Turaga, Senior Member, IEEE 2 2 0 2 n a J 1 ] G L . s c [ 1 v 1 1 1 0 0 . 1 0 2 2 : v i X r a Abstract—Deep neural networks are parametrized by several thousands or millions of parameters, and have shown tremendous success in many classification problems. However, the large number of parameters makes it difficult to integrate these models into edge devices such as smartphones and wearable devices. To address this problem, knowledge distillation (KD) has been widely employed, that uses a pre-trained high capacity network to train a much smaller network, suitable for edge devices. In this paper, for the first time, we study the applicability and challenges of using KD for time-series data for wearable devices. Successful application of KD requires specific choices of data augmentation methods during training. However, it is not yet known if there exists a coherent strategy for choosing an augmentation approach during KD. In this paper, we report the results of a detailed study that compares and contrasts various common choices and some hybrid data augmentation strategies in KD based human activity analysis. Research in this area is often limited as there are not many comprehensive databases available in the public domain from wearable devices. Our study considers databases from small scale publicly available to one derived from a large scale interventional study into human activity and sedentary behavior. We find that the choice of data augmentation techniques during KD have a variable level of impact on end performance, and find that the optimal network choice as well as data augmentation strategies are specific to a dataset at hand. However, we also conclude with a general set of recommendations that can provide a strong baseline performance across databases. Index Terms—Knowledge Distillation, Data Augmentation, time-series, Wearable Sensor Data. I. INTRODUCTION D EEP LEARNING has achieved state-of-the-art perfor- mance in various fields, including computer vision [1], [2], [3], [4], speech recognition [5], [6], and wearable sensors analysis [7], [8]. In general, stacking more layers or increasing the number of learnable parameters causes deep networks to exhibit improved performance [2], [3], [4], [8], [9], [10]. However, this causes the model to become large resulting in additional need for compute and power resources, for training, storage, and deployment. These challenges can hinder the ability to incorporate such models into edge devices. Many studies have explored techniques such as network pruning [11], E. Jeon, A. Shukla and P. Turaga are with the School of Arts, Media and Engineering and School of Electrical, Computer and Energy Engineering, Arizona State University, Tempe, AZ 85281 USA email: (ejeon6@asu.edu; Ankita.Shukla@asu.edu; pturaga@asu.edu). A. Som is with the Center for Vision Technologies Group at SRI Interna- tional, Princeton, NJ 08540 USA email: Anirudh.Som@sri.com K. Hasanaj and M. P. Buman are with the College of Health Solutions, Arizona State University, Phoenix, AZ 85004 USA email: (khasanaj@asu.edu; mbuman@asu.edu). This has been accepted at the IEEE Internet of Things Journal. [12], quantization [12], [13], low-rank factorization [14], and Knowledge Distillation (KD) [15] to compress deep learning models. At the cost of lower classification accuracy, some of these methods help to make the deep learning model smaller and increase the speed of inference on the edge devices. Post- training or fine-tuning strategies can be applied to recover the lost classification performance [12], [13]. On the contrary, KD does not require fine-tuning nor is subjected to any post- training processes. KD is a simple and popular technique that is used to develop smaller and efficient models by distilling the learnt knowl- edge/weights from a larger and more complex model. The smaller and larger models are referred to as student and teacher models, respectively. KD allows the student model to retain the classification performance of the larger teacher model. Recently, different variants of KD have been proposed [16], [17]. These variations rely on different choices of network architectures, teacher models, and various features used to train the student model. Alongside, teacher models trained by early stopping for KD (ESKD) have been explored, which have helped improving the efficacy of KD [18]. However, to the best of our knowledge, there is no previous study that explores the effects, challenges, and benefits of KD for human activity recognition using wearable sensor data. In this paper, we firstly study KD for human activity recog- nition from time-series data collected from wearable sensors. Secondly, we also evaluate the role of data augmentation techniques in KD. This is evaluated by using several time domain data augmentation strategies for training as well as for testing phase. The key highlights and findings from our study are summarized below: • We compare and contrast several KD approaches for time-series data and conclude that EKSD performs better as compared to other techniques. • We perform KD on time-series data with different sizes of teacher and student networks. We corroborate results from previous studies that suggest that the performance of a higher capacity teacher model is not necessarily better. • We study the effects of data augmentation methods on both teacher and student models. We do this to identify which combination of augmentation methods give the most benefit in terms of classification performance. • Our study is evaluated on human activity recognition task and is conducted on a small scale publicly available dataset as well as a large scale dataset. This ensures the observations are reliable irrespective of the dataset sizes. IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 2 Fig. 1. An overview of standard knowledge distillation framework (left) and proposed knowledge distillation with data augmentation method (right). A high capacity network known as teacher is used to guide the learning of a smaller network known as student. A set of augmentations strategies are used to train both the teacher and student networks. The rest of the paper is organized as follows. In Section II, we provide a brief overview of KD techniques as well as data augmentation strategies. In Section III, we present which augmentation methods are used and its effects on time-series data. In Section IV, we describe our experimental results and analysis. In Section V, we discuss our findings and conclusions. II. BACKGROUND 1) Knowledge Distillation: The goal of KD is to supervise a small student network by a large teacher network, such that the student network achieves comparable or improved performance over teacher model. This idea was firstly explored by Buciluˇa et al. [19] followed by several developments like Hinton et al. [15]. The main idea of KD is to use the soft labels which are outputs, soft probabilities, of a trained teacher network and contain more information than just a class label, which is illustrated in Fig. 1. For instance, if two classes have high probabilities for a data, the data has to lie close to a decision boundary between these two classes. Therefore, mimicking these probabilities helps student models to get knowledge of teachers that have been trained with labeled data (hard labels) alone. During training, the loss function L for a student network is defined as: L = (1 − λ)LC + λLK (1) where LC is the standard cross entropy loss, LK is KD loss, and λ is hyper-parameter; 0 < λ < 1. In supervised learning, the error between the output of the softmax layer of a student network and ground-truth label is penalized by the cross-entropy loss: LC = H(sof tmax(as), yg) (2) where H(·) denotes a cross entropy loss function, as is logits of a student (inputs to the final softmax), and yg is a ground truth label. In the process of KD, instead of using peaky probability distributions which may produce less accurate results, Hinton et al. [15] proposed to use probabilities with temperature scaling, i.e., output of a teacher network given by ft = sof tmax(at/τ ) and a student fs = sof tmax(as/τ ) are softened by hyperparameter τ , where τ > 1. The teacher and student try to match these probabilities by a KL-divergence loss: LK = τ 2KL(ft, fs) (3) where KL(·) is the KL-divergence loss function. There has been lots of approaches to improve the per- formance of distillation. Previous methods focus on adding more losses on intermediate layers of a student network to be closer to a teacher [20], [21]. Averaging consecutive student models tends to produce better performance of students [22]. By implementing KD repetitively, the performance of KD is improved, which is called sequential knowledge distillation [23]. Recently, learning procedures for improved efficacy of KD has been presented. Goldblum et al. [24] suggested adver- sarially robust distillation (ARD) loss function by minimiz- ing dependencies between output features of a teacher. The method used perturbed data as adversarial data to train the student network. Interestingly, ARD students even show higher accuracy than their teacher. We adopt augmentation methods to create data which is similar to adversarial data of ARD. Based on ARD, the effect of using adversarial data for KD can be verified, however, which data augmentation is useful for training KD is not well explored. Unlike ARD, to figure out the role of augmentation methods for KD and which method im- proves the performance of KD, we use augmentation methods generating different kinds of transformed data for teachers and students. In detail, by adopting augmentation methods, we can generate various combinations of teachers and students which are trained with the same or different augmentation method. It provides to understand which transformation and combinations can improve the performance of KD. We explain the augmen- tation method for KD in Section III with details. Additionally, KD tends to show an efficacy with transferring information from early stopped model of a teacher, where training strategy is called ESKD [18]. Early stopped teachers produce better students than the standard knowledge distillation (Full KD) using fully-trained teachers. Cho et al. [18] presented the efficacy of ESKD with image datasets. We implement ESKD on time-series data and investigate its efficacy on training with data transformed by various augmentation methods. We explain more details in Section III and discuss the efficiency of ESKD in later sections. IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 3 In general, many studies focus on the structure of networks and adding loss functions to existing framework of KD [25], [26]. However, the performance of most approaches depends on the capacity of student models. Also, availability of suffi- cient training data for teacher and student models can affect to the final result. In this regard, the factors that have an affect on the distillation process need to be systematically explored, especially on time-series data from wearable sensors. 2) Data Augmentation: Data augmentation methods have been used to boost the generalizability of models and avoid over-fitting. They have been used in many applications such as time-series forecasting [27], anomaly detection [28], classifi- cation [8], [29], and so on. There are many data augmentation approaches for time-series data, which can be broadly grouped under two categories [30]. The first category consists of trans- formations in time, frequency, and time-frequency domains [30], [31]. The second group consists of more advanced meth- ods like decomposition [32], model-based [33], and learning- based methods [34], [30]. Time-domain augmentation methods are straightforward and popular. These approaches directly manipulate the orig- inal input time-series data. For example, the original data is transformed directly by injecting Gaussian noise or other perturbations such as step-like trend and spikes. Window cropping or sloping also has been used in time domain transformation, which is similar to computer vision method of cropping samples [35]. Other transformations include window warping that compresses or extends a randomly chosen time range and flipping the signal in time-domain. Additionally, one can use blurring and perturbations in the data points, especially for anomaly detection applications [36]. A few approaches have focused on data augmentation in the frequency domain. Gao et al. [36] proposed perturbations for data augmentation in frequency domain, which improves the performance of anomaly detection by convolutional neural networks. The performance of classification was found to be improved by amplitude adjusted Fourier transform and iterated amplitude adjusted Fourier transform which are transformation meth- ods in frequency domain [37]. Time-frequency augmentation methods have also been recenlty investigated. SpecAugment is a Fourier-transform based method that transforms in Mel- Frequency for speech time-series data [31]. The method was found to improve the performance of speech recognition. In [38], a short Fourier transform is proposed to generate a spectrogram for classification by LSTM neural network. Decomposition-based, model-based, and learning-based methods are used as advanced data augmentation methods. For decomposition, time-series data are disintegrated to create new data [32]. Kegel et al. firstly decomposes the time-series based on trend, seasonality, and residual. Then, finally new time- series data are generated with a deterministic and a stochastic component. Bootstrapping methods on the decomposed resid- uals for generating augmented data was found to help the performance of a forecasting model [39]. Model-based ap- proaches are related to modeling the dynamics, using statistical model [33], mixture models [40], and so on. In [33], model- based method were used to address class imbalance for time- series classification. Learning-based methods are implemented with learning frameworks such as generative adversarial nets (GAN) [34] and reinforcement learning [41]. These methods generate augmented data by pre-trained models and aim to create realistic synthetic data [34], [41]. Finally, augmentation methods can be combined together and applied simultaneously to the data. Combining augmenta- tion methods in time-domain helps to improve performance in classification [42]. However, combining various augmentation methods may results in a large amount of augmented data, increasing training-time, and may not always improve the performance [30]. III. STRATEGIES FOR KNOWLEDGE DISTILLATION WITH DATA AUGMENTATION We would like to investigate strategies for training KD with time-series data and identify augmentation methods for teachers and students that can provide better performance. The strategies include two scenarios on KD. Firstly, we apply augmentation methods only when a student model is trained based on KD with a teacher model trained by the original data. Secondly, augmentation methods are applied not only to students, but also to teacher. When a teacher model is trained from scratch, an augmentation method is used, where the model is to be used as a pre-trained model for distillation. And, when a student is trained on KD, the same/different augmentation methods are used. The set of augmentation approaches on KD are illustrated in Fig. 1, and described in further detail later in this section. Also, we explore the effects of ESKD on time-series data – ESKD uses a teacher which is obtained in the early training process. ESKD generates better students rather than using the fully-trained teachers from Full KD [18]. The strategy is derived from the fact that the accuracy is improved initially. However, the accuracy towards the end of training begins to decrease, which is lower than the earlier accuracy. We adopt early stopped teachers with augmentation methods for our experiments presented in Section IV. Fig. 2. Illustration of different augmentation methods used in our knowledge distillation framework. The original data is shown in blue and the correspond- ing transformed data with data augmentation method is shown in red. IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 4 In order to see effects of augmentation on distillation, we adopt time-domain augmentation methods which are removal, adding noise with Gaussian noise, and shifting. The original pattern, length of the window, and periodical points can be preserved by this transformation. We use transformation meth- ods in time domain so that we can analyze the results from each method, and combinations, more easily. These methods also have been used popularly for training deep learning net- works [30]. We apply combinations of augmentation methods, combined with removal and shifting, and with all methods to a data to see the relationships between each property of datasets for teachers and students of KD. An example of different transformation used for data augmentation is shown in Fig. 2. We describe each of the transforms below: is samples. The values of • Removal: is used to erase amplitude values of se- quential chosen samples to be erased are transformed to the amplitude of the first point. For example, we assume that n samples are chosen as (Xt+1, Xt+2, · · · , Xt+n) and their amplitudes are (At+1, At+2, · · · , At+n) to be erased. At+1 sam- the amplitude of ple Xt+1 and is assigned to (At+1, At+2, · · · , At+n). That is, values (At+1, At+2, · · · , At+n) are mapped to (At+1, At+1, · · · , At+1). The first point and the number of samples to be erased are chosen randomly. The result of removal is shown in Fig. 2 with a green dashed circle. • Noise Injection: To inject noise, we apply Gaussian noise with mean 0 and a random standard deviation. The result of adding noise is shown in Fig. 2 with yellow dashed circles. the first • Shifting: For shifting data, to keep the characteristics such as values of peak points and periodic patterns in the signal, we adopt index shifting and rolling methods to the data for generating new patterns, which means the 100% shifted signal from the original signal by this augmentation corresponds to the original one. For example, assuming the total number of samples are 50 and 10 time-steps (20% of the total number of samples) are chosen to be shifted. The values for amplitude of samples (X1, X2, · · · , X11, · · · X50) are shifting (A41, A42, · · · , A1, · · · , A39, A40) 10 of are (X1, X2, · · · , X11, · · · , X49, X50). of time-steps to be shifted is chosen randomly. Shifting is shown in Fig. 2 with green dashed arrows. (A1, A2, · · · , A11, · · · , A49, A50). time-steps, newly samples number assigned The the By to • Mix1: Applies removal as well as shifting to the same data. • Mix2: Applies removal, Gaussian noise injection, and shifting simultaneously to the data. IV. EXPERIMENTS AND ANALYSIS In this section, we describe datasets, settings, ablations, and results of our experiments. A. Dataset Description We perform experiments on two datasets: GENEActiv [43] and PAMAP2 [44], both of which are wearable sensors based activity datasets. We evaluate multiple teachers and students of various capacities for KD with data augmentation methods. 1) GENEactiv: GENEactiv dataset [43] consists of 29 activities over 150 subjects. The dataset was collected with a GENEactiv sensor which is a light-weight, waterproof, and wrist-worn tri-axial accelerometer. The sampling frequency of the sensors is 100Hz. In our experiments, we used 14 activities which can be categorized as daily activities such as walking, sitting, standing, driving, and so on. Each class has over approximately 900 data samples and the distribution and details for activities are illustrated in Fig. 3. We split the dataset for training and testing with no overlap in subjects. The number of subjects for training and testing are over 130 and 43, respectively. A window size for a sliding window is 500 time-steps or 5 seconds and the process for temporal windows is full-non-overlapping sliding windows. The number of windows for training is approximately 16000 and testing is 6000. Fig. 3. Distribution of GENEactive data across different activities. Each sample has 500 time-steps. 2) PAMAP2: PAMAP2 dataset [44] consists of 18 physical activities for 9 subjects. The 18 activities are categorized as 12 daily activities and 6 optional activities. The dataset was obtained by measurements of heart rate, temperature, accelerometers, gyroscopes, and magnetometers. The sensors were placed on hands, chest, and ankles of the subject. The total number of dimensions in the time-series is 54 and the sampling frequency is 100Hz. To compare with previous methods, in experiments on this dataset, we used leave- one-subject-out combination for validation comparing the ith subject with the ith fold. The input data is in the form of time- series from 40 channels of 4 IMUs and 12 daily activities. To compare with previous methods, the recordings of 4 IMUs are downsampled to 33.3Hz. The 12 action classes are: lying, sitting, standing, walking, running, cycling, nordic walking, ascending stairs, descending stairs, vacuum cleaning, ironing, and rope jumping. Each class and subject are described in Table I. There is missing data for some subjects and the distribution of the dataset is imbalanced. A window size for a sliding window is 100 time-steps or 3 seconds and step size is 22 time-steps or 660 ms for segmenting the sequences, 05001000150020002500300035004000Number of SamplesTreadmill 1mph (0% grade)Treadmill 3mph (0% grade)Treadmill 3mph (5% grade)Treadmill 4mph (0% grade)Treadmill 6mph (0% grade)Treadmill 6mph (5% grade)Seated-fold/stack laundryStand-fidget with hands1 min brush teeth/hairDrive carHard surface walkCarpet with high heels/dress shoesWalk up-stairs 5 floorsWalk down-stairs 5 floorsGENEactiv DatasetIEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 5 TABLE I DETAILS OF PAMAP2 DATASET. THE DATASET CONSISTS OF 12 ACTIVITIES RECORDED FOR 9 SUBJECTS. Lying Sitting Standing Walking Running Cycling Nordic walking Ascending stairs Descending stairs Vacuum cleaning Ironing Rope jumping Sbj.101 407 352 325 333 318 352 302 233 217 343 353 191 Sbj.102 350 335 383 488 135 376 446 253 221 309 866 196 Sbj.103 329 432 307 435 0 0 0 147 218 304 420 0 Sbj.104 344 381 370 479 0 339 412 243 206 299 374 0 Sbj.105 354 402 330 481 369 368 394 207 185 366 496 113 Sbj.106 349 345 365 385 341 306 400 192 162 315 568 0 Sbj.107 383 181 385 506 52 339 430 258 167 322 442 0 Sbj.108 361 342 377 474 246 382 433 168 137 364 496 129 Sbj.109 0 0 0 0 0 0 0 0 0 0 0 92 Sum 2877 2770 2842 3481 1461 2462 2817 1701 1513 2622 3995 721 Nr. of subjects 8 8 8 8 6 7 7 8 8 8 8 6 which allows semi-non-overlapping sliding windows with 78% overlapping [44]. B. Analysis of Distillation For experiments on GENEactiv, we run 200 epochs for each model using SGD with momentum 0.9 and the initial learning rate lr = 0.1. The lr drops by 0.5 after 10 epochs and drops down by 0.1 every [ t 3 ] where t is the total number of epochs. For experiments on PAMAP2, we run 180 epochs for each model using SGD with momentum 0.9 and the initial learning rate lr = 0.05. The lr drops down by 0.2 after 10 epochs and drops down 0.1 every [ t 3 ] where t is the total number of epochs. The results are averaged over 3 runs for both the datasets. To improve the performance, feature engi- neering [45], [46], feature selection, and reducing confusion by combining classes [47] can be applied additionally. However, to focus on the effects of KD which is based on feature- learning [46], feature engineering/selection methods to boost performance are not applied and all classes as specified in Section IV-A are used in the following experiments. 1) Training from scratch to find a Teacher: To find a teacher for KD, we conducted experiments with training from scratch based on two different network architectures: ResNet [1] and WideResNet [48]. These networks have been popularly used in various state-of-the-art studies for KD [16], [17], [24], [18]. We modified and compared the structure having the similar number of trainable parameters. As described in Table II, for training from scratch, WideResNet (WRN) tends to show better performance than ResNet18(k) where k is the dimension of output from the first layer. The increase in accuracy with the dimension of each block is similar to the basic ResNet. 2) Setting hyperparameters for KD: For setting hyper- parameters in KD, we conducted several experiments with different temperature τ as well as lambda λ. We investigated distillation with different hyperparameters as well. We set WRN16-3 as a teacher network [18] and WRN16-1 as a student network, which is shown in Fig. 4. For temperature τ , in general, τ ∈ {3, 4, 5} are used [18]. High temperature mitigated the peakiness of teachers and helped to make the signal to be softened. In our experiments, according to the results from different τ , high temperature did not effectively help to increase the accuracy. When we used τ = 4, the results were better than other choices for both datasets with Full KD Fig. 4. Effect of hyperparmeters τ and λ on the performance of Full KD and ESKD approaches. The results are reported on GENEactive dataset with WRN16-3 and WRN16-1 networks for teacher and student models respectively. and ESKD [18]. For λ = 0.7 and 0.99, we obtained the best results with Full KD and ESKD for GENEactiv and PAMAP2, respectively. 3) Analyzing Distillation with different size of Models: To analyze distillation with different size of models, WRN16-k and WRN28-k were used as teacher networks having different capacity and structures in depth and width k. WRN16-1 and WRN28-1 were used as student networks, respectively. As mentioned in the previous section, in general, a higher capacity network trained from scratch shows better accuracy for WRN16 and WRN28. However, as shown in Fig. 5, in IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 6 TABLE II ACCURACY FOR VARIOUS MODELS TRAINED FROM SCRATCH ON GENEACTIV Model ResNet18(8) ResNet18(16) ResNet18(24) ResNet18(32) ResNet18(48) ResNet18(64) # Parameters 62,182 244,158 545,942 967,534 2,170,142 3,851,982 Accuracy (%) 63.75±0.42 65.84±0.69 66.47±0.21 66.33±0.12 68.13±0.22 68.17±0.21 Model WRN16-1 WRN16-2 WRN16-3 WRN16-4 WRN16-6 WRN16-8 # Parameters 61,374 240,318 536,254 949,438 2,127,550 3,774,654 Accuracy (%) 67.66±0.37 67.84±0.36 68.89±0.56 69.00±0.22 70.04±0.05 69.02±0.15 Model WRN28-1 - WRN28-2 WRN28-3 WRN28-4 WRN28-6 # Parameters 126,782 - 500,158 1,119,550 1,985,214 4,455,358 Accuracy (%) 68.63±0.48 - 69.15±0.24 69.23±0.27 69.29±0.51 70.99±0.44 TABLE V ACCURACY (%) FOR RELATED METHODS ON GENEACTIV DATASET WITH 7 CLASSES Method WRN16-1 WRN16-3 WRN16-8 ESKD (WRN16-3) ESKD (WRN16-8) Full KD (WRN16-3) Full KD (WRN16-8) SVM [49] Choi et al. [50] Window length 1000 89.29±0.32 89.53±0.15 89.31±0.21 89.88±0.07 (89.74) 89.58±0.13 (89.68) 89.84±0.21 (88.95) 89.36±0.06 (88.74) 86.29 89.43 500 86.83±0.15 87.95±0.25 87.29±0.17 88.16±0.15 (88.30) 87.47±0.11 (87.75) 87.05±0.19 (86.02) 86.38±0.06 (85.08) 85.86 87.86 Fig. 5. Results of distillation from different teacher models of WRN16-k and WRN28-k on GENEactiv dataset. The higher capacity of teachers does not always increase the accuracy of students. TABLE VI ACCURACY FOR RELATED METHODS ON PAMAP2 DATASET TABLE III ACCURACY FOR VARIOUS MODELS ON GENEACTIV DATASET Student WRN16-1 (ESKD) WRN16-1 (Full KD) Teacher WRN16-2 WRN16-3 WRN16-4 WRN16-6 WRN16-8 WRN16-3 WRN16-8 Teacher Acc. (%) 69.06 69.99 69.80 70.24 70.19 69.68 69.28 Student Acc. (%) 69.34±0.36 69.49±0.22 69.37±0.31 67.93±0.13 68.62±0.33 68.62±0.22 68.68±0.17 most of the cases, the results from WRN16-k shows better than the results of WRN28-k which has larger width. And the accuracy with teachers of WRN16-3 is higher than the one with teachers having larger width. Therefore, a teacher of higher capacity is not always guaranteed to generate a student whose accuracy is better. 4) Knowledge Distillation based on Fully Iterated and Early Stopped Models: We performed additional experiments TABLE IV ACCURACY FOR VARIOUS MODELS ON PAMAP2 DATASET Student WRN16-1 (ESKD) WRN16-1 (Full KD) Teacher WRN16-2 WRN16-3 WRN16-4 WRN16-6 WRN16-8 WRN16-3 WRN16-8 Teacher Acc. (%) 84.86 85.67 85.23 85.51 85.17 81.52 81.69 Student Acc. (%) 86.18±2.44 86.38±2.25 85.95±2.27 86.37±2.35 85.11±2.46 84.31±2.24 83.70±2.52 Method WRN16-1 WRN16-3 WRN16-8 ESKD (WRN16-3) ESKD (WRN16-8) Full KD (WRN16-3) Full KD (WRN16-8) Chen and Xue [51] Ha et al.[52] Ha and Choi [53] Kwapisz [54] Catal et al. [55] Kim et al.[56] Accuracy (%) 82.81±2.51 84.18±2.28 83.39±2.26 86.38±2.25 (85.67) 85.11±2.46 (85.17) 84.31±2.24 (81.52) 83.70±2.52 (81.69) 83.06 73.79 74.21 71.27 85.25 81.57 with WRN16-k which gives the best results. Table III and Table IV give detailed results for GENEactiv and PAMAP2, respectively. Compared to training from scratch, although the student capacity from KD is much lower, the accuracy is higher. For instance, for the result of GENEactiv with WRN16- 8 by training from scratch, the accuracy is 69.02% and the number of trainable parameters is 3 million in Table III. The number of parameters for WRN16-1 as a student for KD is 61 thousand which is approximately 1.6% of 3 million. However, the accuracy of a student with WRN16-2 teacher from ESKD is 69.34% which is higher than the result of training from scratch with WRN16-8. It shows a model can be compressed with conserved or improved accuracy by KD. Also, we tested with 7 classes on GENEactiv dataset which were used by the 123468WRN16/28-k (k: width)66.567.067.568.068.569.069.570.070.5Accuracy (%)KD with WRN16/WRN28WRN16-kWRN28-kIEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 7 method in [50]. This work used over 50 subjects for testing set. Students of KD were WRN16-1 and trained with τ = 4 and λ = 0.7. As shown in Table V where brackets denote the structure of teachers and their accuracy, ESKD from WRN16-3 teacher shows the best accuracy for 7 classes, which is higher than results of models trained from scratch, Full KD, and previous methods [49], [50]. In most of the cases, students are even better than their teacher. In various sets of GENEactiv having different number of classes and window length, ESKD shows better performance than Full KD. In Table IV, the best accuracy on PAMAP2 is 86.38% from ESKD with teacher of WRN16-3, which is higher than results from Full KD. The result is even better than previous methods [57], which are described in Table VI where brackets denote the structure of teachers and their accuracy. Therefore, KD allows model compression and improves the accuracy across datasets. And ESKD tends to show better performance compared to Full KD. Also, the higher capacity models as teachers does not always generate better performing student models. C. Effect of augmentation on student model training To understand distillation effects based on the various capacity of teachers and augmentation methods, WRN16-1, WRN16-3, and WRN16-8 are selected as “Small”, “Medium”, and “Large” models, respectively. ESKD is used for this experiment which tends to show better performance than the Full KD and requires three-fourths of the total number of epochs for training [18]. In order to find augmentation methods impacting KD on students for training, we first trained a teacher from scratch with the original datasets. Secondly, we trained students from the pre-trained teacher with augmentation methods which have different properties including removal, adding noise, shifting, Mix1, and Mix2. For experiments on GENEactiv, for removal, the number of samples to be removed is less than 50% of the total number of samples. The first point and the exact number of samples to be erased are chosen randomly. To add noise, the value for standard deviation of Gaussian noise is chosen uniformly at random between 0 and 0.2. For shifting, the number of time-steps to be shifted is less than 50% of the total number of samples. For Mix1 and Mix2, the same parameters are applied. For experiments on PAMAP2, the number of samples for removal is less than 10% of the total number of samples and standard deviation of Gaussian noise for adding noise is less than 0.1. The parameter for shifting is less than 50% of the total number of samples. The same parameters of each method are applied for Mix1 and Mix2. The length of the window for PAMAP2 is only 100 which is 3 seconds and downsampled from 100Hz data. Compared to GENEactiv whose window size is 500 time-steps or 5 seconds, for PAMAP2, a small transformation can affect the result very prominently. Therefore, lower values are applied to PAMAP2. The parameters for these augmentation methods and the sensor data for PAMAP2 to be transformed are randomly chosen. These conditions for applying augmentation methods are used in the following experiments as well. TABLE VII ACCURACY (%) OF TRAINING FROM SCRATCH ON WRN16-1 WITH DIFFERENT AUGMENTATION METHODS Method Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) Dataset GENEactiv 68.60±0.23 69.20±0.32 67.60±0.36 68.69±0.22 69.31±0.96 67.89±0.11 PAMAP2 82.81±2.51 83.34±2.41 82.80±2.66 83.91±2.18 83.59±2.37 83.64±2.76 Fig. 6. The validation accuracy for training from scratch and Full KD. WRN16-1 is used for training from scratch. For Full KD, WRN16-3 is a teacher network and WRN16-1 is a student network. R, N, S, M1, and M2 in the legend are removal, adding noise, shifting, Mix1, and Mix2, respectively. 1) Analyzing augmentation methods on training from scratch and KD: The accuracy of training scratch with dif- ferent augmentation methods on WRN16-1 is presented in Table VII. Most of the accuracies from augmentation methods, except adding noise which can alter peaky points and change gradients, are higher than the accuracy obtained by learning with the original data. Compared to other methods, adding noise may influence classification between similar activities such as walking, which is included in both datasets as detailed sub-categories. The validation accuracy of scratch and Full KD learning on GENEactiv dataset is presented in Fig. 6. Training from scratch with the original data shows higher accuracy than KD with original data in very early stages before 25 epochs. However, KD shows better accuracy than the models trained from scratch after 40 epochs. KD with augmentation tends to perform better in accuracy than models trained from scratch and KD learning with the original data alone. That is, data augmentation can help to boost the generalization ability of student models for KD. Mix1 shows the highest accuracy among the results. The highest accuracies are seen in early stages, which are less than 120 epochs for all methods, where 120 epochs is less than three-fourths of the total number of epochs. On closer inspection, we find that the best accuracies are actually seen in less than 20 epochs for training from scratch and Full KD, less than 60 epochs for shifting, Mix1, and Mix2, and less than 120 epochs for adding noise, respec- tively. This implies that not only early stopped teachers but 020406080100120Epochs45505560657075Accuracy (%)Scratch/Full KD with AugmentationScratchFullFull-RFull-NFull-SFull-M1Full-M2IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 8 TABLE VIII ACCURACY (%) OF KD FROM VARIANTS OF TEACHER CAPACITY AND AUGMENTATION METHODS ON GENEACTIV (λ = 0.7) TABLE XI ACCURACY (%) OF KD FROM VARIANTS OF TEACHER CAPACITY AND AUGMENTATION METHODS ON PAMAP2 (λ = 0.99) Method Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) Small 68.87 69.71±0.31 69.80±0.34 69.26±0.08 70.63±0.19 70.56±0.57 69.27±0.31 Teacher Medium 69.99 69.61±0.17 70.23±0.41 69.12±0.19 70.43±0.89 71.35±0.20 69.51±0.28 Large 70.19 68.62±0.33 70.28±0.68 69.38±0.39 70.00±0.20 70.22±0.10 69.62±0.21 Method Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) Small 85.42 86.37±2.35 84.66±2.67 84.77±2.65 86.08±2.42 84.93±2.71 82.94±2.76 Teacher Medium 85.67 86.38±2.25 85.70±2.40 85.21±2.41 86.65±2.13 85.88±2.28 83.94±2.70 Large 85.17 85.11±2.46 84.81±2.52 85.05±2.40 85.53±2.28 84.73±2.54 83.28±2.50 TABLE IX ACCURACY (%) OF KD FROM VARIANTS OF TEACHER CAPACITY AND AUGMENTATION METHODS ON GENEACTIV (λ = 0.99) Method Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) Small 68.87 69.44±0.19 69.48±0.22 69.99±0.14 70.96±0.10 70.40±0.27 70.56±0.23 Teacher Medium 69.99 67.80±0.36 69.75±0.40 70.20±0.06 70.42±0.06 70.07±0.38 69.88±0.16 Large 70.19 68.67±0.20 70.01±0.27 70.12±0.14 70.16±0.24 69.36±0.16 69.71±0.30 also early stopped students are able to perform better than fully iterated models. In training based on KD with augmentation methods, the accuracy goes up in early stages, however, the accuracy suffers towards to the end of training. These trends on KD are similar to the previous ESKD study [18]. For the following experiments, we restrict our analyses to ESKD. 2) Analyzing Augmentation Methods on Distillation: The accuracy of each augmentation method with KD is summa- rized in Table VIII and IX for GENEactiv and Table X and XI for PAMAP2. The results were obtained from small-sized students of ESKD. The gray colored cells of these tables are the best accuracy for the augmentation method among the different capacity teachers of KD. When a higher λ is used, distillation from teachers is improved, and the best results are obtained when the teacher capacity is smaller. Also, the best performance of students, when learning with augmentation methods and the original data, is achieved with similar teacher capacities. For example, for GENEactiv with λ = 0.7, the best results are generated from various capacity of teachers. But, with λ = 0.99, the best results tend to be seen with smaller capacity of teachers. Even though the evaluation protocol for PAMAP2 is leave-one-subject-out with TABLE X ACCURACY (%) OF KD FROM VARIANTS OF TEACHER CAPACITY AND AUGMENTATION METHODS ON PAMAP2 (λ = 0.7) an imbalanced distribution of data, with λ = 0.7, the best results are obtained from larger capacity of teachers as well. Furthermore, results from both datasets verify that larger and more accurate teachers do not always result in better students. Also, the best result from shifting is seen at the same capacity of the teacher with the original data. It might be because shifting includes the same time-series ‘shapes’ as the original data. The method for shifting is simple but is an effectively helpful method for training KD. For all teachers on PAMAP2 with λ = 0.99, the accuracies from training by shifting are even higher than other combinations. Compared to previous methods [57] with PAMAP2, the result by shifting outperforms others. Furthermore, although the student network of KD has the same number of parameters of the network trained from scratch (WRN16-1), the accuracy is much higher than the latter one; the result of Mix1 from GENEactiv and shifting from PAMAP2 by the medium teacher is approximately 2.7% points and 3.8% points better than the result from original data by training from scratch, respectively. These accuracies are even better than the results of their teachers. It also verifies that KD with an augmentation method including shifting has benefits to obtain improved results. TABLE XII p-VALUE AND (ACCURACY (%), STANDARD DEVIATION) FOR TRAINING FROM SCRATCH AND KD ON GENEACTIV DATASET Scratch Original (68.60±0.23) KD (Teacher: Medium) Original (ESKD) (69.61±0.17) Original (Full) (68.62±0.22) Removal (70.23±0.41) Noise (69.12±0.19) Shift (70.43±0.89) Mix1(R+S) (71.35±0.20) Mix2(R+N+S) (69.51±0.28) p-value 0.030 0.045 0.006 0.012 0.025 0.073 0.055 Method Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) Small 85.42 84.75±2.64 85.16±2.46 84.96±2.59 85.21±2.21 85.54±2.51 85.17±2.39 Teacher Medium 85.67 84.47±2.32 85.51±2.27 85.52±2.26 85.45±2.19 85.60±2.19 85.27±2.33 Large 85.17 84.90±2.38 85.02±2.47 84.85±2.43 85.66±2.26 84.71±2.53 83.76±2.77 To investigate the difference in performance with a model trained from scratch and KD with augmentation methods, statistical analysis was conducted by calculating p-value from a t-test with a confidence level of 95%. Table XII and XIII show averaged accuracy, standard deviation, and calculated p- value for WRN16-1 trained from scratch with original training set and various student models of WRN16-1 trained with KD IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 9 TABLE XIII p-VALUE AND (ACCURACY (%), STANDARD DEVIATION) FOR TRAINING FROM SCRATCH AND KD ON PAMAP2 DATASET Scratch Original (82.81±2.51) KD (Teacher: Medium) Original (ESKD) (84.47±2.32) Original (Full) (84.31±2.24) Removal (85.51±2.27) Noise (85.52±2.26) Shift (85.45±2.19) Mix1(R+S) (85.60±2.19) Mix2(R+N+S) (85.27±2.33) p-value 0.0298 0.0007 0.0008 0.0002 0.0034 0.0024 0.0013 and augmentation. That is, student models in KD have the same structure of the model trained from scratch and teachers for KD are WRN16-3 (τ = 4, λ = 0.7). For GENEactiv, in five out of the seven cases, the calculated p-values are less than 0.05. Thus, the results in the table show statistically- significant difference between training from scratch and KD. in all cases, p-values are less than 0.05. For PAMAP2, This also represents statistically-significant difference between training from scratch and KD. Therefore, we can conclude that KD training with augmentation methods, which shows better results in classification accuracy, performs significantly different from training from scratch, at a confidence level of 95%. TABLE XIV ECE (%) OF TRAINING FROM SCRATCH AND KD ON GENEACTIV DATASET Scratch Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) ECE KD (Teacher: Medium) Original (ESKD) 3.22 Removal 3.56 Noise 3.45 3.24 Shift 3.72 Mix1(R+S) 3.67 Mix2(R+N+S) ECE 2.96 2.90 2.85 2.78 2.79 2.86 TABLE XV ECE (%) OF TRAINING FROM SCRATCH AND KD ON PAMAP2 DATASET Scratch Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) ECE KD (Teacher: Medium) Original (ESKD) 2.28 Removal 3.64 Noise 5.83 2.87 Shift 4.39 Mix1(R+S) 5.55 Mix2(R+N+S) ECE 2.16 3.09 3.01 2.22 2.96 4.17 Finally, the expected calibration error (ECE) [58] is calcu- lated to measure the confidence of performance for models trained from scratch and KD (τ = 4, λ = 0.7) with aug- mentation methods. As shown in Table XIV and XV, in all cases, ECE values for KD are lower than when models are trained from scratch, indicating that models trained with KD have higher reliability. Also, results of KD including shifting are lower than results from other augmentation methods. This additionally verifies that KD improves the performance and shifting helps to get improved models. TABLE XVI THE LOSS VALUE (10−2) FOR KD (TEACHER: MEDIUM) FROM VARIOUS METHODS ON GENEACTIV Method (λ=0.7) Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) CE Train 3.774 3.340 11.687 2.416 5.475 17.420 KD Train 0.617 0.406 1.172 0.437 0.475 1.337 KD Test 1.478 1.246 1.358 1.119 1.108 1.338 TABLE XVII THE LOSS VALUE (10−2) FOR KD (TEACHER: MEDIUM) FROM VARIOUS METHODS ON PAMAP2 (SUBJECT 101) Method (λ=0.7) Original Removal Noise Shift Mix1(R+S) Mix2(R+N+S) CE Train 0.832 1.237 1.066 0.468 1.267 1.853 KD Train 0.156 0.146 0.138 0.129 0.150 0.177 KD Test 1.783 1.038 1.284 1.962 0.895 1.065 3) Analyzing training for KD with augmentation methods: The loss values of each method, for the medium-sized teacher, are shown in Table XVI and XVII. The loss values were obtained from the final epoch while training student models based on Full KD. As shown in these tables, for both cross entropy and KD loss values, training with shifting-based data augmentation results in lower loss, compared to other augmentation strategies and the original model. The loss value for noise augmentation is higher than the values of shifting. On the other hand, the KD loss value for Mix1 is higher than the values for removal and shifting. However, the training loss is for these two methods and its value of testing is lower. Compared to other methods, Mix2 shows higher loss for training, which may be because this method generates more complicated patterns. However, the testing KD loss value of Mix2 is lower than the value of original and adding noise. These findings imply that the data of original and shifting have very similar patterns. And data based on Mix1 and Mix2 are not simply trainable data for distillation, however, these methods have an effect of preventing a student from over- fitting or degradation for classification. The contrast of results from GENEactiv between each method is more prominent than the one from PAMAP2. This is due to the fact that smaller parameters for augmentation are applied to PAMAP2. Also, the dataset is more challenging to train on, due to imbalanced data and different channels in sensor data. D. Analysis of Teacher and Student Models with a Variant Properties of Training Set To discuss properties of training set for teacher and student models, we use the same parameter (τ = 4, λ = 0.7) in this IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 10 experiment on two datasets. In this section, we try to train a medium teacher and a small student by training set having the same or different properties to take into account relationships between teachers and students. Testing set is not transformed or modified. The medium teacher is chosen because the teacher showed good performance in our prior experiments discussed in previous sections. Further, distillation from a medium model to a small model is an preferable approach [18]. Also, we analyze which augmentation method is effective to achieve higher accuracy. We use adding noise, shifting, and Mix1 methods which transform data differently. TABLE XVIII ACCURACY (%) OF TRAINING FROM SCRATCH ON WRN16-3 WITH DIFFERENT AUGMENTATION METHODS Dataset GENEactiv GENEactiv (Top-1) PAMAP2 PAMAP2 (Top-1) Original 69.53±0.40 Noise 68.59±0.05 Shift 72.08±0.20 Mix1(R+S) 71.64±0.26 69.99 68.68 72.48 72.17 84.65±2.28 83.08±2.51 82.54±2.42 82.39±2.62 85.67 85.31 84.38 84.09 To obtain a medium teacher model, the model is trained from scratch with augmentation methods. These results are shown in Table XVIII. For GENEactiv, shifting based data augmentation gives the best performance. However, for PAMAP2, original data achieves the best performance. Mix1 shows slightly lower accuracy than shifting. In these experi- ments, the student model is trained using the teacher model that achieves best performance over several trials. We also evaluated different combinations of data augmen- tation strategies for teacher-student network pairs. A pair is obtained by using one or no data augmentation strategy to train the teacher network by training from scratch, and the student network is trained by ESKD under different, same, or no augmentation strategy. The results are shown in Fig. 7. We found that KD with the same data augmentation strategy for training teachers and students may not be the right choice to get the best performance. When a teacher is trained by is trained by Mix1 which showed shifting and a student good performance as a student in the previous sections, the results are better than other combinations for both datasets. Also, when a student is learned by Mix1 including shifting transform, in general, the performance are also good for all teachers. It implies that the method chosen for training a student is more important than choosing a teacher; KD with a medium teacher trained by the original data and a student trained with shift or Mix1 outperforms other combinations. Using the same strategy for training data for teachers and students does not always present the best performance. When the training set for students is more complicated than the set for teachers, the performance in accuracy tends to be better. That is, applying a transformation method to students can help to increase the accuracy. It also verifies that better teachers do not always lead to increased accuracy of students. Even if the accuracies from these combinations of a teacher and student are lower than models trained from scratch by WRN16-3, the number of parameters for the student is only about 11% of the Fig. 7. The results for students trained by different combinations of training sets for teachers and students. The teacher and student both are learned by augmentation methods. WRN16-3 (medium) and WRN16-1 (small) are teacher and student networks, respectively. one for WRN16-3. Therefore, the results still are good when considering both performance and computation. E. Analysis of Student Models with Different Data Augmen- tation Strategies for Training and Testing Set In this section, we study the effect of students on KD from various augmentation methods for training and testing, while a teacher is trained with the original dataset. We use the same parameter (τ = 4, λ = 0.7) and ESKD for this experiment on two datasets. A teacher is selected with a medium model trained by the original data. We use adding noise, shifting and Mix1 methods which transform data differently. After training the teacher network on original data, a student network is trained with different data augmentation strategies and is evaluated on test data transformed with different data augmentation strategies. The results are illustrated in Fig. training student networks 8. For GENEactiv, most often, with Mix1 show better performance on different testing sets. However, if the testing set is affected by adding noise, training students with adding noise and Mix2 shows much better performance than training with shifting and Mix1. From the results on PAMAP2, in most of the cases, training students IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 11 time for testing, as shown in Table XIX. WRN16-1 as a student trained by ESKD with Mix1 augmentation achieves the best accuracy, 71.35%, where the model takes the least amount of time on both GPU and CPU. The results on CPU reiterate the reason why model compression is required for many applications, especially on edge devices, wearables, and mobile devices, which have limited computational and power resources and are generally implemented in real time with only CPU. The gap in performance would be higher if an edge device had lower computational resources. TABLE XIX PROCESSING TIME OF VARIOUS MODELS FOR GENEACTIVE DATASET Model (WRN16-k) k=1 k=1 (ESKD) k=1 (ESKD+Mix1) k=3 k=6 k=8 Acc. (%) 67.66 69.61 71.35 68.89 70.04 69.02 Total GPU (sec) Avg. GPU (ms) Total CPU (sec) Avg. CPU (ms) 15.226 2.6644 16.655 2.8920 16.426 16.663 16.885 2.8524 2.8934 2.9320 21.333 33.409 46.030 3.7044 5.8012 7.9928 V. CONCLUSION In this paper, we studied many relevant aspects of knowl- edge distillation (KD) for wearable sensor data as applied to human activity analysis. We conducted experiments with different sizes of teacher networks to evaluate their effect on KD performance. We show that a high capacity teacher network does not necessarily ensure better performance of a student network. We further showed that training with augmentation methods and early stopping for KD (ESKD) is effective when dealing with time-series data. We also establish that the choice of augmentation strategies has more of an impact on the student network training as opposed to the teacher network. In most cases, KD training with the Mix1 (Removal+Shifting) data augmentation strategy for students showed robust performance. Further, we also conclude that a single augmentation strategy is not conclusively better all the time. Therefore, we recommend using a combination of augmentation methods for training KD in general. In summary, our findings provide a comprehensive understanding of KD and data augmentation strategies for time-series data from wearable devices of human activity. These conclusions can be used as a general set of recommendations to establish a strong baseline performance on new datasets and new applications. ACKNOWLEDGMENT This research was funded by NIH R01GM135927, as part of the Joint DMS/NIGMS Initiative to Support Research at the Interface of the Biological and Mathematical Sciences, and by NSF CAREER grant 1452163. REFERENCES [1] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778. [2] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition, 2017, pp. 4700–4708. Fig. 8. Effect on classification performance of student network with different augmentation methods for training and testing sets. WRN16-3 (medium) and WRN16-1 (small) are teacher and student networks, respectively. with Mix1 shows better performance to many different testing set. However, when the testing set is augmented by adding noise, training with original data shows the best performance. This is likely attributable to the window size, which has about a hundred samples, and the dataset includes the information of 4 kinds of IMUs. Therefore, injecting noise, which can affect peaky points and change gradients, creates difficulties for classification. Also, these issue can affect the both training and testing data. Thus, if the target data includes noise, training set and augmentation methods have to be considered along with the length of the window and intricate signal shapes within the windows. F. Analysis of Testing Time Here, we compare the evaluation time for various models on the GENEactiv dataset. We conducted the test on a desktop with a 3.50 GHz CPU (Intel® Xeon(R) CPU E5-1650 v3), 48 GB memory, and NVIDIA TITAN Xp (3840 NVIDIA® CUDA® cores and 12 GB memory) graphic card. We used a batch size of 1 and approximately 6000 data samples for testing. Four different models were trained from scratch with WRN16-k (k=1, 3, 6, and 8). To test with ESKD and Mix1, WRN16-3 was used as a teacher and WRN16-1 was used for student network. As expected, larger models take more IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 12 [3] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2005, pp. 886–893. [25] T. Furlanello, Z. Lipton, M. Tschannen, L. Itti, and A. Anandkumar, the International “Born again neural networks,” in Proceedings of Conference on Machine Learning, 2018, pp. 1607–1616. [4] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004. [5] O. Abdel-Hamid, A.-r. Mohamed, H. Jiang, L. Deng, G. Penn, and D. Yu, “Convolutional neural networks for speech recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 22, no. 10, pp. 1533–1545, 2014. [6] W. Xiong, L. Wu, F. Alleva, J. Droppo, X. Huang, and A. Stolcke, “The microsoft 2017 conversational speech recognition system,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, 2018, pp. 5934–5938. [7] S. Wan, L. Qi, X. Xu, C. Tong, and Z. Gu, “Deep learning models for real-time human activity recognition with smartphones,” Mobile Networks and Applications, vol. 25, no. 2, pp. 743–755, 2020. [8] H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Deep learning for time series classification: a review,” Data Mining and Knowledge Discovery, vol. 33, no. 4, pp. 917–963, 2019. [9] A. Khan, A. Sohail, U. Zahoora, and A. S. Qureshi, “A survey of the recent architectures of deep convolutional neural networks,” Artificial Intelligence Review, vol. 53, no. 8, pp. 5455–5516, 2020. [10] M. Gil-Mart´ın, R. San-Segundo, F. Fernandez-Martinez, and J. Ferreiros- L´opez, “Improving physical activity recognition using a new deep learning architecture and post-processing techniques,” Engineering Ap- plications of Artificial Intelligence, vol. 92, p. 103679, 2020. [11] P. Molchanov, S. Tyree, T. Karras, T. Aila, and J. Kautz, “Pruning convo- lutional neural networks for resource efficient inference,” in Proceedings of the International Conference on Learning Representations, 2017. [12] S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,” in Proceedings of the International Conference on Learning Representations, 2016. [13] J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng, “Quantized convolutional the IEEE neural networks for mobile devices,” in Proceedings of Conference on Computer Vision and Pattern Recognition, 2016, pp. 4820–4828. [14] C. Tai, T. Xiao, Y. Zhang, X. Wang et al., “Convolutional neural net- works with low-rank regularization,” in Proceedings of the International Conference on Learning Representations, 2016. [15] G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” in Proceedings of the International Conference on Neural Information Processing Systems Deep Learning and Representation Learning Workshop, 2015. [16] J. Yim, D. Joo, J. Bae, and J. Kim, “A gift from knowledge distillation: Fast optimization, network minimization and transfer learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 4133–4141. [17] B. Heo, M. Lee, S. Yun, and J. Y. Choi, “Knowledge distillation with adversarial samples supporting decision boundary,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, 2019, pp. 3771– 3778. [18] J. H. Cho and B. Hariharan, “On the efficacy of knowledge distillation,” the IEEE International Conference on Computer in Proceedings of Vision, 2019, pp. 4794–4802. [19] C. Buciluˇa, R. Caruana, and A. Niculescu-Mizil, “Model compression,” the ACM SIGKDD International Conference on in Proceedings of Knowledge Discovery and Data Mining, 2006, pp. 535–541. [20] S. Zagoruyko and N. Komodakis, “Paying more attention to attention: Improving the performance of convolutional neural networks via at- tention transfer,” in Proceedings of the International Conference on Learning Representations, 2017. [21] F. Tung and G. Mori, “Similarity-preserving knowledge distillation,” in Proceedings of the IEEE International Conference on Computer Vision, 2019, pp. 1365–1374. [22] A. Tarvainen and H. Valpola, “Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,” in Proceedings of the International Conference on Neural Information Processing Systems, 2017, pp. 1195–1204. [23] Y. Zhang, T. Xiang, T. M. Hospedales, and H. Lu, “Deep mutual learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 4320–4328. [24] M. Goldblum, L. Fowl, S. Feizi, and T. Goldstein, “Adversarially robust distillation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 04, 2020, pp. 3996–4003. [26] C. Yang, L. Xie, S. Qiao, and A. L. Yuille, “Training deep neural net- works in generations: A more tolerant teacher educates better students,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, 2019, pp. 5628–5635. [27] Z. Han, J. Zhao, H. Leung, K. F. Ma, and W. Wang, “A review of deep learning models for time series prediction,” IEEE Sensors Journal, vol. 21, no. 6, 2019. [28] R. Chalapathy and S. Chawla, “Deep learning for anomaly detection: A survey,” arXiv preprint arXiv:1901.03407, 2019. [29] A. Le Guennec, S. Malinowski, and R. Tavenard, “Data augmentation for time series classification using convolutional neural networks,” in ECML/PKDD workshop on advanced analytics and learning on temporal data, 2016. [30] Q. Wen, L. Sun, X. Song, J. Gao, X. Wang, and H. Xu, “Time series data augmentation for deep learning: A survey,” arXiv preprint arXiv:2002.12478, 2020. [31] D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” in Proceedings of the Interspeech, 2019, pp. 2613–2617. [32] L. Kegel, M. Hahmann, and W. Lehner, “Feature-based comparison the International and generation of time series,” in Proceedings of Conference on Scientific and Statistical Database Management, 2018, pp. 1–12. [33] H. Cao, V. Y. Tan, and J. Z. Pang, “A parsimonious mixture of gaussian trees model for oversampling in imbalanced and multimodal time-series classification,” IEEE Transactions on Neural Networks and Learning Systems, vol. 25, no. 12, pp. 2226–2239, 2014. [34] C. Esteban, S. L. Hyland, and G. R¨atsch, “Real-valued (medical) time series generation with recurrent conditional gans,” arXiv preprint arXiv:1706.02633, 2017. [35] X. Cui, V. Goel, and B. Kingsbury, “Data augmentation for deep neural network acoustic modeling,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, no. 9, pp. 1469–1477, 2015. [36] J. Gao, X. Song, Q. Wen, P. Wang, L. Sun, and H. Xu, “Robusttad: Ro- bust time series anomaly detection via decomposition and convolutional neural networks,” arXiv preprint arXiv:2002.09545, 2020. [37] K. T. L. Eileen, Y. Kuah, K.-H. Leo, S. Sanei, E. Chew, and L. Zhao, “Surrogate rehabilitative time series data for image-based deep learning,” in Proceedings of the European Signal Processing Conference, 2019, pp. 1–5. [38] O. Steven Eyobu and D. S. Han, “Feature representation and data augmentation for human activity classification based on wearable imu sensor data using a deep lstm neural network,” Sensors, vol. 18, no. 9, p. 2892, 2018. [39] C. Bergmeir, R. J. Hyndman, and J. M. Ben´ıtez, “Bagging exponential smoothing methods using stl decomposition and box–cox transforma- tion,” International journal of forecasting, vol. 32, no. 2, pp. 303–312, 2016. [40] Y. Kang, R. J. Hyndman, and F. Li, “Gratis: Generating time series with diverse and controllable characteristics,” Statistical Analysis and Data Mining: The ASA Data Science Journal, vol. 13, no. 4, pp. 354–376, 2020. [41] E. D. Cubuk, B. Zoph, D. Mane, V. Vasudevan, and Q. V. Le, “Autoaug- ment: Learning augmentation strategies from data,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 113–123. [42] T. T. Um, F. M. Pfister, D. Pichler, S. Endo, M. Lang, S. Hirche, U. Fietzek, and D. Kuli´c, “Data augmentation of wearable sensor data for parkinson’s disease monitoring using convolutional neural networks,” in Proceedings of the 19th ACM International Conference on Multimodal Interaction, 2017, pp. 216–220. [43] Q. Wang, S. Lohit, M. J. Toledo, M. P. Buman, and P. Turaga, “A statistical estimation framework for energy expenditure of physical activities from a wrist-worn accelerometer,” in Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 2016, 2016, pp. 2631–2635. [44] A. Reiss and D. Stricker, “Introducing a new benchmarked dataset for activity monitoring,” in Proceedings of the International Symposium on Wearable Computers, 2012, pp. 108–109. [45] A. Zheng and A. Casari, Feature engineering for machine learning: principles and techniques for data scientists. O’Reilly Media, Inc., 2018. IEEE INTERNET OF THINGS JOURNAL, VOL. 0, NO. 0, JANUARY 2022 13 [46] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new perspectives,” IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 8, pp. 1798–1828, 2013. [47] A. Dutta, O. Ma, M. P. Buman, and D. W. Bliss, “Learning approach for classification of geneactiv accelerometer data for unique activity identification,” in 2016 IEEE 13th International Conference on Wearable and Implantable Body Sensor Networks (BSN). IEEE, 2016, pp. 359– 364. [48] S. Zagoruyko and N. Komodakis, “Wide residual networks,” in Proceed- ings of the British Machine Vision Conference, 2016. [49] C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning, vol. 20, no. 3, pp. 273–297, 1995. [50] H. Choi, Q. Wang, M. Toledo, P. Turaga, M. Buman, and A. Srivastava, “Temporal alignment improves feature quality: an experiment on activity the IEEE recognition with accelerometer data,” in Proceedings of Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 349–357. [51] Y. Chen and Y. Xue, “A deep learning approach to human activity recognition based on single accelerometer,” in Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics, 2015, pp. 1488–1492. [52] S. Ha, J.-M. Yun, and S. Choi, “Multi-modal convolutional neural net- works for activity recognition,” in Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics, 2015, pp. 3017–3022. [53] S. Ha and S. Choi, “Convolutional neural networks for human activity recognition using multiple accelerometer and gyroscope sensors,” in Proceedings of the International Joint Conference on Neural Networks, 2016, pp. 381–388. [54] J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition us- ing cell phone accelerometers,” ACM SigKDD Explorations Newsletter, vol. 12, no. 2, pp. 74–82, 2011. [55] C. Catal, S. Tufekci, E. Pirmit, and G. Kocabag, “On the use of ensemble of classifiers for accelerometer-based activity recognition,” Applied Soft Computing, vol. 37, pp. 1018–1022, 2015. [56] H.-J. Kim, M. Kim, S.-J. Lee, and Y. S. Choi, “An analysis of eating activities for automatic food type recognition,” in Proceedings of the Asia Pacific Signal and Information Processing Association Annual Summit and Conference, 2012, pp. 1–5. [57] A. Jordao, A. C. Nazare Jr, J. Sena, and W. R. Schwartz, “Human activity recognition based on wearable sensor data: A standardization of the state-of-the-art,” arXiv preprint arXiv:1806.05226, 2018. [58] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in Proceedings of the International Confer- ence on Machine Learning, 2017, pp. 1321–1330. Eun Som Jeon received the B.E. and M.E. de- grees in Electronics and Electrical Engineering from Dongguk University, Seoul, Korea in 2014 and 2016, respectively. She worked in Korea Telecom (Institute of Convergence Technology), Seoul, Korea. She is currently pursuing the Ph.D. degree in Computer Engineering (Electrical Engineering) with Geometric Media Laboratory, Arizona State University, Tempe, AZ, USA. Her current research interests include time-series and image data analysis, human behavior analysis, deep learning, and artificial analysis. Anirudh Som is an Advanced Computer Scientist in the Center for Vision Technologies group at SRI International. He received his M.S. and Ph.D. degrees in Electrical Engineering from Arizona State University in 2016 and 2020 respectively, prior to which he received his B.Tech. degree in Electron- ics and Communication Engineering from GITAM University in India. His research interests are in the fields of machine learning, computer vision, human movement analysis, human behavior analysis and dynamical system analysis. Ankita Shukla is a postdoctoral researcher at Ari- zona State University. She received her PhD and Masters degrees in Electronics and Communication from IIIT-Delhi, India in 2020 and 2014 respectively. Her research interest are in the field of machine learning, computer vision, time-series data analysis and geometric methods. Kristina Hasanaj is a graduate research associate at Arizona State University. She earned her B.S. in Exercise Science (Kinesiology concentration) and M.A. in Exercise Physiology from Central Michigan University. She is currently pursuing her doctoral degree through the Nursing and Healthcare Innova- tion Ph.D. program at Arizona State University. Her research interests are focused around behaviors in the 24-hour day (sleep, sedentary behavior, physical activity) and the use of mobile health and wearable technologies in clinical and health related settings. Matthew P. Buman , PhD is an associate professor in the College of Health Solutions at Arizona State University. His research interests reflect the dynamic interplay of behaviors in the 24-hour day, including sleep, sedentary behavior, and physical activity. His work focuses on the measurement of these behav- iors using wearable technologies, interventions that singly or in combination target these behaviors, and the environments that impact these behaviors. Pavan Turaga , PhD is an associate professor in the School of Arts, Media and Engineering at Arizona State University. He received a bachelor’s degree in electronics and communication engineering from the Indian Institute of Technology Guwahati, India, in 2004, and a master’s and doctorate in electrical en- gineering from the University of Maryland, College Park in 2007 and 2009, respectively. His research interests include computer vision and computational imaging with applications in activity analysis, dy- namic scene analysis, and time-series data analysis with geometric methods. 

synthetic_cpt

"The_Promise_and_Challenge_of_Large_Language_Models_for_Knowledge_Engineering_Insights_from_a_Hackat(...TRUNCATED)

"Using Large Language Models for Knowledge\nEngineering (LLMKE): A Case Study on Wikidata\n\nBohui Z(...TRUNCATED)

synthetic_cpt

Medical_Image_Synthesis_via_Fine-Grained_Image-Text_Alignment_and_Anatomy-Pathology_Prompting.pdf

"4\n2\n0\n2\n\nr\na\n\nM\n1\n1\n\n]\n\nV\nC\n.\ns\nc\n[\n\n1\nv\n5\n3\n8\n6\n0\n.\n3\n0\n4\n2\n:\nv\(...TRUNCATED)

synthetic_cpt

"Increasing_Diversity_While_Maintaining_Accuracy_Text_Data_Generation_with_Large_Language_Models_and(...TRUNCATED)

"4\n2\n0\n2\n\ng\nu\nA\n5\n1\n\n]\n\nG\nL\n.\ns\nc\n[\n\n1\nv\n6\n5\n0\n8\n0\n.\n8\n0\n4\n2\n:\nv\ni(...TRUNCATED)

synthetic_cpt

Search_Query_Spell_Correction_with_Weak_Supervision_in_E-commerce.pdf

"Spelling Correction with Denoising Transformer\n\nAlex Kuznetsov\nHubSpot, Inc.\nDublin, Ireland\na(...TRUNCATED)

synthetic_cpt

Smaller_Weaker_Yet_Better_Training_LLM_Reasoners_via_Compute-Optimal_Sampling.pdf

"0\n1\n0\n2\n\nr\np\nA\n1\n1\n\n]\n\nG\nM\n.\nh\nt\na\nm\n\n[\n\n1\nv\n2\n5\n8\n1\n.\n4\n0\n0\n1\n:\(...TRUNCATED)

synthetic_cpt

"Full-dose_PET_Synthesis_from_Low-dose_PET_Using_High-efficiency_Diffusion_Denoising_Probabilistic_M(...TRUNCATED)

"8\n1\n0\n2\n\nr\na\n\nM\n3\n2\n\n]\nS\nD\n.\nh\nt\na\nm\n\n[\n\n3\nv\n7\n7\n2\n7\n0\n.\n5\n0\n7\n1\(...TRUNCATED)

synthetic_cpt

Towards_Robust_Evaluation_of_Unlearning_in_LLMs_via_Data_Transformations.pdf

"Abhinav Joshi♣\nSriram Vema⋄\n\nTowards Robust Evaluation of Unlearning in LLMs via Data\nTrans(...TRUNCATED)

synthetic_cpt

ASVD_Activation-aware_Singular_Value_Decomposition_for_Compressing_Large_Language_Models.pdf

"4\n2\n0\n2\n\nt\nc\nO\n9\n2\n\n]\nL\nC\n.\ns\nc\n[\n\n4\nv\n1\n2\n8\n5\n0\n.\n2\n1\n3\n2\n:\nv\ni\n(...TRUNCATED)