_id
stringlengths
36
36
text
stringlengths
200
328k
label
stringclasses
5 values
0e0b6de7-d0f5-4d3e-bcc6-80d7819e2ebd
Representations of dataset relationship. The term indicator matrix has been proposed in [1]}, [2]} for primary PK-FK-joinable or equi-joinable tables. They share similar ideas to ours in terms of using binary matrices to indicate the matches between rows from different tables. In our vision, the semantics of indicator matrices are richer. Besides row-level matching of inner joins as discussed in [1]}, [2]}, we have also covered how users choose to combine the datasets, i.e., the table-level relationships such as outer join, left join, and union as shown in Table REF . Moreover, the declarative representations in our vision include also mapping matrix and redundancy matrix. They preserve another two fundamental aspects of relationships among sources, i.e., source-to-target schema-level mapping and data redundancy. It then leads to new research challenges of LA rewriting, as discussed in Sec. REF , and also broadens future research directions, e.g., federated learning as discussed in Sec. REF .
w
c5f86565-de73-4e77-aa21-c63afde405d4
In this work, we focus on data integration challenges over different sources for machine learning tasks. We proposed three types of matrices for schema mapping, data matching, and data redundancy. With these simple, easy-to-use representations, we have shown the new opportunities for model factorization and federated learning, which enable new systems such as our data science platform Amalur. Our vision is that the proposed representations become one of the first steps towards bridging the recent advances in machine learning with the well-studied classical data integration field. We hope that our proposed solution and outlined challenges would inspire more interesting future work.
d
7f316b15-934a-445e-bd2e-8026ced40ff4
The appeal of Machine Learning (ML) lies much in its ability, given sample data, to automatically discover complex patterns that work well also for out-of-sample predictions mullainathan2017machine, requiring minimal data domain expertise. However, the learning part in ML typically amounts to training a model to minimize a cost function on a known and limited sample. Thus, we could state that the outcome of the learning process is driven by the selection of the training dataset, by the definition of the cost function, and by the choice of the algorithm used to optimise the latter. While in the majority of applications these choices are quite straightforward, there are three widely agreed conditions in which conscious and non-trivial interventions are necessary: Class Imbalance (CI) ali2015classification, Confounding Bias (CB) adeli2019bias and Unfair Classification (UC) mehrabi2019survey.
i
56a1c97a-4f69-4efb-8b3a-aab490cdff16
CI occurs when one class in the training set is overrepresented with respect to the others. This imbalance is critical because errors on the majority class can overwhelm and mask those on the minority class, leading to models that predict almost uniquely the majority category, at the expense of the others. In the most severe cases, the process of errors' minimization could favor a model that exclusively predicts the overrepresented class. CI is widely documented in biomedical applications as well as in fault detection, where diseases, pathological conditions and faults often present very skewed class distributions ali2015classification. It can also be easily spotted in machine vision lin2017focal, e.g. when pixels are classified as a rarely occurring event (i.e. presence of an object) or background (the majority class).
i
1545e2dc-c774-4e42-83ea-c33ea000c317
CB occurs when the distribution of the target classes in the training set is biased with respect to a non-target variable, called confounder, which instead is not related to the expected real distribution of the classes. This bias can mislead the learning process, leading to a model that bases its predictions on the associations between the input data and the confounder. The misleading effect of unwanted biases is a well known problem in causal inference bareinboim2012controlling,kuroki2014measurement, but it has only recently gained attention in the ML context. In fact, much of earlier effort in ML focused on achieving generalization in predictive performance, which has been historically treated as distinct from causal investigations kreif2019machine. However, in many ML applications, the objective is causal in nature, even if not specifically framed as such castro2020causality. For instance, the paper by Caruana et al. caruana2015intelligible describes a model to predict the probability of death for pneumonia, trained on medical records of patients who have previously had pneumonia. Counter-intuitively, the model found that asthma lowers the risk of death, while it is known to be a severe condition in subjects with pneumonia. This misleading effect occurred because the patients with asthma in the training set received more care by the hospital system. In this example, the association learned from the dataset was correct, but clearly the aim of this application was to find only causal relations useful to prioritize care for patients with pneumonia.
i
8998870c-efa2-4cb2-bdff-2463f201af27
If causal reasoning is only recently spreading in the ML context, the necessity to understand the reasons behind predictions has been discussed in the new field of Explainable Artificial Intelligence (XAI). In this context, Ribeiro et al. ribeiro2016should developed a prediction explanation technique and showed how it can be used to check whether a model's predictions are due to spurious correlations in the training set. For example, they trained an image classifier to distinguish between wolves and huskies and showed that the predictions of the model were based on the presence of snow in the background of the picture, which was more likely in the wolves images. In another XAI-related paper, Ross et al. ross2017right introduced the concept of a model right for the right reasons and developed a method for producing and explaining multiple classifiers that are accurate for qualitatively different reasons, leaving the responsibility to choose the right one to a domain expert.
i
097a99c5-dc86-4894-bc43-a55db5a9a650
A special case of CB is UC, that occurs when the bias involves a sensitive feature and thus the resulting model is unfair towards a protected group. Despite being well known that UC derives from the bias problem (a recent review lists 23 different types of biases causing unfairness mehrabi2019survey), given the increasing number of ML applications taking life-changing decisions ajunwa2016hiring, fairness issues attracted more attention than CB, forming a distinct branch of studies. The societal impact of UC pushed the development of solutions based on the imposition of fairness properties to the classifier, while the recent works to address CB usually focus on the causes of the problem, trying to avoid learning from the confounder variable.
i
88d52719-8f93-4f30-88b9-d6b731fe9db0
Summarizing, CI, CB and UC are conditions in which, without any active corrective intervention, the minimization of the prediction error on the training set may produce a model that does not satisfy the purposes for which it was developed: either because it is not informative with respect to a class (the CI case), or because it has learned a wrong pattern and thus it is not generalizable (the CB case) or finally because the predictions are discriminatory towards a protected group (the UC case).
i
34c10db5-362b-406a-984b-4ae6e04bb8b0
Usually, these three conditions are studied separately and the corrections developed for them are based on different strategies. Instead, in this paper we show that they are different expressions of unbalance: Class Imbalance occurs when the target classes are unbalanced with respect to each other, while when they are unbalanced with respect to another variable we talk about Confounding Bias. Finally, if this bias involves protected attributes we refer to it as Unfair Classifications.
i
6b6885e3-ce67-44b2-a278-2969576a23ef
Given that unbalance is the common cause of all the three conditions, we introduce a new unified corrective loss function to address Fairness, Bias and Imbalance problems (FBI-loss). We experimentally test the performances of our loss on six benchmarks comprising synthetic and real-world data, affected by each issue under consideration in this work, confronting it with other solutions specifically tailored for each problem.
i
d2e1e660-b6c8-4b3f-8ac5-ff1847ec193d
We discuss a novel unifying perspective over CI, CB and UC problems, highlighting similarities and differences when considered as problems of learning with adaptively penalized losses. We introduce a new approach to tackle CI, CB and UC conditions, that relies solely on a modification of the training loss. Our approach applies generically to any classification problem (though for the sake of conciseness it is described for the binary case alone), and does not pose any restriction on the nature of the learning model. We provide an experimental assessment of the proposed method on both synthetic and real-world data, confronting our FBI-loss with state-of-the-art approaches from CI, CB and UC. As a consequence of the previous point, this work shows an in-depth performance comparison of various methods to deal with different forms of unbalance, exploring many configurations of data type, task complexity and unbalance degree.
i
0dcf0694-9118-474a-9b35-0b9ba88cab10
In this section we briefly overview the research on CI, CB and UC problems. Synthesizing the contents of the following paragraphs, the strategies to deal with CI are mainly aimed at re-balancing the data or the error committed on them, while to address CB, the objective is focused on removing the dependency of the classification results or of the training data from the confounder; finally, to cope with UC, all approaches start from one definition of fairness and try to build a classifier that satisfies that definition.
i
5acc99e0-1ce2-4e09-87df-f3e197f83623
In this analysis, we compare the performances of our loss with related approaches from literature using synthetic datasets with varying degrees of unbalance and different task complexities. Unbalance is computed as the ratio between the number of over-represented instances in the training set and its total size. As stated above, the inverse measure of task complexity is instead given by the parameter \(\theta _Y\) .
m
d2396395-4a54-44a2-828c-55f6889f4009
Besides \(H^*\) , that constitutes a baseline reference, all the other approaches to which we compare our loss (see Table REF ) depend on one or more hyperparameters. To perform a fair comparison, when testing a model we let its hyperparameters vary for every combination of unbalances and complexities under consideration, and we select the configuration that maximizes simultaneously the performances on OverG and UnderG. The ranges of hyperparameters spanned are reported in the last column of Table REF .
m
5f24d1f4-c9c8-4007-898d-80ac649720bc
With this second analysis we extend the comparison made on synthetic data to real-world applications. This means that the complexity is intrinsically given by the dataset under examination. Thus, to compare the various approaches we simply study the trend of the performance metric with respect to unbalance, making sure to use the same training and validation sets for all the methods tested on the same dataset.
m
86d431c4-1661-4095-ab79-523e79a3f389
CI: since the data used for this evaluation consist in small 1-dimensional feature vectors, for all methods under consideration we adopt a simple architecture comprising a single-layer perceptron with 16 hidden neurons. CB: in this application, data are 3D images; therefore, for the FBI loss we use a lighter version of AlexNet krizhevsky2012imagenet (adapted to the 3D case) composed by 10 convolutional layers and 3 linear layers with a total of about 2 million trainable parameters. We analogously build the feature extractors of BR-NN and IFR-VAE with convolutional and linear layers. UC: Given the simplicity of both data and task, for all the approaches we use a single-layer neural network with 8 neurons.
m
fc44ba0d-c8dc-43f9-892c-0d1733e3d924
The results of this analysis are more straightforward to visualise and interpret with respect to the ones described in the synthetic analysis. A perfectly correcting model should show high and stable performances across the different unbalances on both UnderG and OverG.
r
fe18571b-77ed-4180-bc42-54bf07013527
Class Imbalance, Confounding Bias and Fairness issues can severely undermine the usefulness of Machine Learning applications. In this work, we describe them from a unified perspective and we provide a unified formalization highlighting that they all arise from a different type of unbalance. We reviewed the most recent solutions developed for each problem and suggested instead a unified cost-corrected loss (called FBI-loss) to address all these well-known learning problems at once. The main idea inherited by the Focal loss, specifically developed for Class imbalance, is to counteract unbalance by taking into account also the complexity of the classification problem. Our FBI-loss is straightforward to implement, requires tuning of a single hyperparameter, and it is model-agnostic and general enough to be used with any classifier and for any kind of data. Furthermore, its only hyperparameter directly depends on the complexity of the task, which makes the tuning process easier for a data-domain expert. When tested on both synthetic and real-world benchmarks, the FBI-loss shows performances that are competitive with those from relevant related approaches from literature and specifically tailored to tackle one single issue. Despite this preliminary formulation is suitable only for binary classification applications, the performance and versatility of the proposed loss should encourage future works in the direction of unified solutions.
d
da14dc93-7a88-4286-a572-bac409eb4111
Language representation models include the Bidirectional Encoder Representations from Transformers (BERT) , underlie the most successful systems currently on the leaderboard hosted by the Allen Institute for a variety of question answering and commonsense reasoning benchmarks .
i
07e7fbbf-ec16-425f-8c21-dc6ed5995005
RQ1: If we give a set of low-value items and a set of high-value items to LRMs, can they correctly identify them by separating high-value items from the low-value items? RQ2: Can these pre-trained LRMs, after fine-tuned on structure similar text, i.e., question answering problems, or bet questions, make rational choices?
i
fc2f5c22-e008-450c-9f24-82324fb627b5
An influential model is Bidirectional Encoder Representations from Transformers (BERT) , which used a novel attention mechanism to obtain rich `pre-trained' representations of language from a large corpus of text, followed by task-specific `fine-tuning' that allowed it to get state-of-the-art performance (at the time) on specific tasks, such as question answering, without requiring full training from scratch. In experiments, BERT was found to obtain a score of 80.5% on the General Language Understanding Evaluation (GLUE) benchmark , an improvement of 7.7% over the previous best performing model. Similarly, it achieves an F1 score of 93.1% on the Stanford Question Answering Dataset (SQuAD) v1.1 benchmark and 83.1% on SQuAD v2.0 , with improvements of 1.5% and 5.1% compared to the previous leading model, respectively. Moreover, the pre-trained BERT model was released publicly and is amenable to fine-tuning on other tasks. We draw on this capability in the proposed work.
w
d0096a11-1a76-44d4-9155-9dffd975f9bd
A more mature version is the Robustly Optimized BERT Pretraining Approach (RoBERTa) . RoBERTa is structurally the same as BERT. However, RoBERTa improves the training process on several points, such as a bigger batch size, more extended sequence, and longer training. RoBERTa also removes the next sentence prediction objective and introduces the dynamic masking strategy. Therefore, compared to BERT on their published GLUE and SQuAD performance, RoBERTa shows significant improvements and obtains new state-of-the-art results on 4 of the GLUE tasks. Like BERT, the pre-trained RoBERTa model was released and is amenable to fine-tune.
w
27c9487c-6788-45e0-b187-4d646cd6431d
The third model is the single pre-trained QA model (UnifiedQA) . UnifiedQA breaks the format restrictions on Question Answering(QA) tasks by introducing new ways to unify different formats: Extractive, Yes/No, Multiple-Choice, and Abstractive. The backbone structure of UnifiedQA is the Text-to-Text Transfer Transformer (T5) , which is also based on Transformer , just like BERT and RoBERTa. UnifiedQA achieves new state-of-the-art results on ten factoid and commonsense QA benchmarks by unifying different formats.
w
544c15d8-7b1d-4b1d-a6fc-597ad4b48730
Modern Language Representation Models (LRMs), based on transformer neural networks, have rapidly exceeded the previous state-of-the-art on a range of NLP tasks, including question answering, text summarization, and information extraction. In this article, we addressed the question of whether, and how, such LRMs can be adapted for (approximately) rational decision-making and preference elicitation. In the cognitive science literature, preference elicitation is often evaluated using bets. Given the advanced, near human-like performance, of LRMs on language-based Artificial Intelligence problems, we formulated a set of research hypotheses to specifically test whether: (i) LRMs can distinguish between high- and low-value items, even when the items were not seen in the training set, and after we control for the format of the questions, (ii) LRMs can make, or be taught to make, (approximately rational) bets in a generalizable manner, including when it has been trained on one type (or `modality') of bet, but is evaluated on another.
d
9864669e-d932-44de-8f9e-b98451018508
We construct a novel benchmark to empirically test these hypotheses using three established transformer-based LRMs. Our first set of results show that, while LRMs can distinguish between high- and low-value items, they only do so after we control for the format of the questions. This is despite the question being expressed in relatively simple language. Our second set of results is similar: LRMs can only make bets (whether posed using the same, or different, modality as the training set) once they have been trained on some pre-existing bet questions. We find, furthermore, that changing the modality of the bet typically leads to a noticeable drop in performance, but is still much higher than random. Thus, while the models do seem to be generalizing, they are not doing so (yet) at a sufficiently abstract level.
d
c121c667-51e3-439a-b00d-e74c32b99029
There are many promising avenues for future research. Earlier, we alluded to formulating and testing hypotheses on why larger and more recently released LRMs (such as UnifiedQA) seem to be performing better on bet-style decision-making problems compared to the slightly older generation of LRMs. More interestingly, our experiments have only probed the surface of these models' decision-making abilities, since our benchmark tests decision-making when the number of outcomes are limited and equi-probable, and the putative value difference between the pair of items is extreme. It remains to be tested whether even the newer models would be able to effectively generalize to more complex decisions and question-structures. Considering that there are infinitely many decision-making modalities, what general methodology is required for the LRMs to achieve human-level decision-making performance, short of training on many different modalities? We note that, in the real world, decisions are not framed so precisely or explicitly, and the modalities used (if any) are unknown a priori. To be applicable in such situations, therefore, the LRM needs to be taught decision-making as a fundamental ability, as opposed to some form of brute-force training on ever larger corpora.
d
f39df81d-3b1e-4162-9ea7-91e777d4f95e
Finally, another direction that could be explored is the use of generative models, such as the Generative Pre-trained Transformer 3 (GPT-3) model, for decision making. Such models have recently yielded promising results on zero-shot learning problems, and have even been shown to exhibit human-like creativity. They may be more amenable to decision making in an open-ended setting when no option is given (or can be computed in advance), but an evaluation to that effect has not been conducted yet.
d
9aa33985-083b-4319-b2c1-d37a5e95c5c5
Fractals are ubiquitous in nature [1]} and increasingly prevalent in artificial structures [2]}, [3]}. Coastlines, rivers, and trees all exhibit self-similar structures—accurate replicas of themselves within themselves. Fractals are also prevalent across length scales in animals and cellular structures: respiratory and vascular systems, brains [4]}, and DNA [5]} all exploit space filling fractal networks. At the largest known length scale, the universe itself consists of a fractal clustering of spiraling galaxies [6]}. However, the ways in which fractals may be useful in robotics is mostly unexplored.
i
382fe902-7c85-40a8-9cdd-fc4c2b7d8d64
One desirable property of fractals derives from their ability to represent infinite complexity within a compact representation: a simple rule applied to itself, recursively. [1]} exploited this phenomenon in robotics by evolving branching kinematic chains in simulation and then building physical instances of the most promising designs [2]}. The resulting robots yielded an order of magnitude more physical elements than any other simulated robots at that time: Whereas [3]} and [4]} evolved simulated robots composed of no more than 14 and 50 parts, respectively, [1]} utilized Lindenmayer-systems [6]} to produce branching structures of up to 350 parts.
i
113f7495-f59b-4f39-b963-1aa5892343d8
The ubiquity of self-similar forms in living systems suggests that fractals may confer adaptive benefits in addition to reduced descriptive complexity, which could be useful in artificial systems such as robots. One such potential use of fractals in robotics is that self similar structure can, in some cases, result in self similar behavior. Branching arteries and capillaries maximize efficient blood flow at the macro and micro scales, respectively. In the gecko, nanoscale spatulae branch from microscale setae, which in turn branch from toes, which in turn branch from feet, and together maximize the probability of adhesion of spatulae, setae, toes, and ultimately the animal itself to vertical surfaces. However, with the exception of artificial gecko feet [1]} and Moravec and Easudes' hypothetical trillion-fingered Bush Robot [2]}, this particular property of fractals remains unexplored in robotics. Here, we explore whether self similar structure can confer self similar behavior in modular robots.
i
b55b663b-a6be-46ff-86ac-cd4e87d96ca5
Conventional robots contain numerous smaller components (legs, wheels, end effectors, servomotors, sensors, battery packs) but these parts are highly specialized, interdependent, and incapable of independent or self-similar behavior. Modular robots differ in that they consist of repeated robotic elements that are to some degree self-sufficient, able to behave and survive on their own. They often autonomously detach, move about each other, and reattach to reconfigure the robot's overall geometry [1]}, [2]}, [3]}. But robot modules tend to be indivisible cubes or other simple geometries that do not reprise the shape of the whole robot, unless the robot's overall structure is a cube of cubes. In all cases, the target behavior of a module is, by design, different from that of the whole robot. <FIGURE>
i
14cc5a80-35e0-4dca-bfd5-17a0edeeb5e6
Swarms of hundreds or thousands of individual robots [1]} have loosely coupled to form amorphous phototatic aggregates [2]} or moved within a shared shell [3]}, however these robots have largely been restricted to cylindrical geometries and 2D interaction in-plane. Most robot swarms comprise rigid-bodied electromechanical robots which, in order to be produced in large quantities, are designed to be as simple as possible. A consequence of this is that each robot may be more or less behaviorally static (functionless) in isolation [2]}, [3]}.
i
9d4cbcd6-8899-4ba6-aab6-23da24aea704
Here, we test whether self-similar forms can facilitate the evolution of scale invariant behavior as follows. Shapes of basal robots (e.g. Fig REF a) are evolved in a voxel-based simulation; each design can be composed with copies of itself, using the same pattern, indefinitely to realize self-similar form at a range of size scales. However, if the desired behavior is lost and needs to be relearned de novo at each level of recursion, then training such a system to operate on more than one size scale becomes computationally infeasible. Thus, an evolutionary algorithm is here employed to find behaviorally scale-invariant fractal robots: they must demonstrate similar behavior, normalized for scale, at three different size scales. The results demonstrate that robots with this property can be automatically designed and in some cases, manufactured in reality.
i
e3ff7993-2a5d-4f38-9e23-7d9af64eae78
In the first attempts to transfer simulated scale invariant behavior to a physical system, computer-generated mold designs were 3D printed and then used to rapidly fabricate identical, hollow elastomeric modules that can be de/pressurized to induce volumetric actuation [1]}. Modules behaved in isolation and together when attached to form a fractal soft robot.
i
756c34d9-55bd-4b73-9f95-0a97247f1fc6
Scale invariant fractal design principles could also provide novel solutions to challenges facing in vitro bioengineering. One such challenge is the size limit of biobots without vascular systems in which diffusion does not bring sufficient oxygen and other key metabolites beyond approximately 1 mm. As proof of principle, we show that living xenobots, motile biological machines built from amphibian stem cells, can be mechanically joined forming permanent multi-individual subunits that do not need vacular systems because they always keep their cells close to an interface with a nutrient medium. These subunits could then be stacked and multiplexed to form increasingly large fractal biobots. <FIGURE><FIGURE><FIGURE><FIGURE><FIGURE>
i
9d1e7d9d-3881-4787-8c59-da001a25e461
In this paper, we achieved scale invariant robot behavior using fractals. The results demonstrate that, under certain conditions, the nonrandom behavior of a single robot can be preserved when connected with copies of itself fractally. However, this depends on the morphology of the basal robot; most robot morphologies do not exhibit scale invariant behavior. We found that evolutionary algorithmsIt is unlikely that evolutionary algorithms are uniquely capable of designing fractal robots with scale invariant behavior. Other computational search methods (e.g. reinforcement learning [1]} and simulated annealing [2]}) are capable of producing self similar robot forms, which in principle could be modified to generate scale invariant behavior. can automatically discover morphologies that facilitate scale invariant robot behavior in simulation.
d
c32348f2-87f0-412e-8425-724c1524e892
Simulated fractal robots were transferred to reality in two systems. Using pneumatically actuated silicone bladders, two centimeters across, self similar structure was readily assembled, and the physical robot exhibited scale invariant behavior. However, the amount of locomotion was limited and the direction did not always match the simulated prediction. The micrometer length scale biobots investigated here exhibited spontaneous attachment and scale invariant behavior up to supermillimeter length scales. However, self similar structure was more challenging to produce out of plane.
d
8850e099-887a-4817-85f9-ff409719f70b
isprime begins with the sieve of Eratosthenes followed by trial division of all primes up to the square root of the maximum value, M, in the array [1]}. The sieve begins with an incrementing sequence of \([2, 3, 4, 5, ..., \emph {M}]\) . Then it crosses-off multiples of primes. Two, for example, is prime so all multiples of two are crossed-off given that they share a common factor. This leaves \([2, 3, 5, 7, 9, ..., \emph {M}]\) . The next prime is three: it is marked prime then its multiples are crossed-off leaving \([2, 3, 5, 7, 11, ..., \emph {M}]\) . The process repeats until all prime numbers have been visited.
i
91770fa7-4fe3-476b-8a76-17fa0c953ead
Next, isprime utilizes trial division of all primes returned by the sieve. isprime divides each element of the input array by all primes which are less than that element. The input element is marked prime when none of the smaller primes divides into it evenly. This algorithm is simple, effective, and accurate, but is not optimized for any category of inputs.
i
b723c2b0-1e0b-461c-8bea-46b981117614
This research introduces isprime_fast, a replacement for isprime. isprime_fast's inputs and outputs are identical; the only difference lies in the techniques used to determine primality. isprime_fast specializes for four classifications of inputs: (1) large scalar integers, (2) medium scalar integers, (3) small scalar integers, and (4) arrays.
i
6076173b-2600-4b29-9023-242865518bdc
The largest improvement of isprime_fast over MATLAB's isprime is for 64-bit integers where isprime_fast runs 250x to 3000x faster. The modern Miller—Rabin algorithm for 64-bit numbers is tremendously fast compared to isprime's classic prime number sieve followed by prime factor division. One reason for this great contrast is because isprime's implementation predates MATLAB's support for 64-bit integers: the int64 and uint64 data types were not supported until 2004 [1]}. Before then, MATLAB only supported the double-precision data type which has a maximum odd value of \(2^{53} - 1\) . This is significant because isprime completes in one second for 53-bit numbers but its duration balloons to over 30 seconds for 64-bit integers.
d
d30d9c0a-1dda-45c9-a894-3b3d81644628
The largest area for improvement of isprime_fast is for scalar inputs in the 40- to 48-bit range. Figure REF shows a dip in performance for input scalars in this range: isprime_fast only executes two to three times faster than isprime. isprime_fast uses a 2/3/5/7-wheel sieve for inputs in this range. For 48-bit integers, this wheel is over 3 million elements long and therefore requires equally as many divisions to determine primality. An alternative algorithm likely could remedy this performance depression.
d
7836fefd-4727-46b5-bd88-8ee3081c6600
MATLAB's isprime implementation is stable, effective, and depended on by many mathematics enthusiasts and researchers. isprime provides fair performance for small to medium size scalars and short arrays, but less than optimal performance otherwise. isprime_fast, a proposed replacement for isprime, improves upon isprime for all 264 scalar inputs, ranging from 2.4x to 2950x faster. isprime_fast also surpasses isprime for all arrays which were tested, normally ranging from 2x to 1400x faster. isprime_fast efficiently creates wheel sieves, implements the Miller—Rabin test using modular arithmetic strategies, minimizes divisions, and heavily relies on vectorized operations. isprime_fast employs the power of the MATLAB language: it requires no arbitrary-precision arithmetic, C/C++ source code, or external libraries. Using the techniques detailed here, MATLAB's esteemed isprime function can be improved considerably.
d
ec9cbab3-7786-47e7-998e-d0a7f8fae06e
Despite the fact that Wikipedia exists in 287 languages, the existing content is unevenly distributed. The content of the most under-resourced Wikipedias is maintained by a limited number of editors – they cannot curate the same volume of articles as the editors of large Wikipedia language-specific communities. It is therefore of the utmost social and cultural interests to address languages for which native speakers have only access to an impoverished Wikipedia. In this paper, we propose an automatic approach to generate textual summaries that can be used as a starting point for the editors of the involved Wikipedias. We propose an end-to-end trainable model that generates a textual summary given a set of KB triples as input. We apply our model on two languages that have a severe lack of both editors and articles on Wikipedia: Esperanto is an easily acquired artificially created language which makes it less data needy and a more suitable starting point for exploring the challenges of this task. Arabic is a morphologically rich language that is much more challenging to work, mainly due to its significantly larger vocabulary. As shown in Table REF both Arabic and Esperanto suffer a severe lack of content and active editors compared to the English Wikipedia which is currently the biggest one in terms of number of articles.
i
fc0c37da-0cec-483b-bf98-f77a0b6f0c8d
Our research is mostly related to previous work on adapting the general encoder-decoder framework for the generation of Wikipedia summaries [1]}, [2]}, [3]}. Nonetheless, all these approaches focus on task of biographies generation, and only in English – the language with the most language resources and knowledge bases available. In contrast with these works, we explore the generation of sentences in an open-domain, multilingual context. The model from [1]} takes the Wikipedia infobox as an input, while [2]} uses a sequence of slot-value pairs extracted from Wikidata. Both models are only able to generate single-subject relationships. In our model the input triples go beyond the single-subject relationships of a Wikipedia infobox or a Wikidata page about a specific item (Section ). Similarly to our approach, the model proposed by [3]} accepts a set of triples as input, however, it leverages instance-type-related information from DBpedia in order to generate text that addresses rare or unseen entities. Our solution is much broader since it does not rely on the assumption that unseen triples will adopt the same pattern of properties and entities' instance types pairs as the ones that have been used for training. To this end, we use copy actions over the labels of entities in the input triples. This relates to previous works in machine translation which deals with rare or unseen word problem for translating names and numbers in text. [7]} propose a model that generates positional placeholders pointing to some words in source sentence and copy it to target sentence (copy actions). [8]} introduce separate trainable modules for copy actions to adapt to highly variable input sequences, for text summarisation. For text generation from tables, [1]} extend positional copy actions to copy values from fields in the given table. For Question Generation, [10]} use a placeholder for the subject entity in the question to generalise to unseen entities.
i
016f87bc-c976-4592-ae42-0962a1fc46c2
We evaluate our approach by measuring how close our synthesised summaries can be to actual summaries in Wikipedia against two other baselines of different natures: a language model, and an information retrieval template-based solution. Our model substantially outperforms all the baselines in all evaluation metrics in both Esperanto and Arabic. In this work we present the following contributions: i) We investigate the task of generating textual summaries from Wikidata triples in underserved Wikipedia languages across multiple domains, and ii) We use an end-to-end model with copy actions adapted to this task. Our datasets, results, and experiments are available at: https://github.com/pvougiou/Wikidata2Wikipedia. <TABLE><TABLE>
i
b6f0e981-1d4c-45b1-8eb5-59eee0ceaa58
In this paper, we show that with the adaptation of the encoder-decoder neural network architecture for the generation of summaries we are able to overcome the challenges introduced by working with underserved languages. This is achieved by leveraging data from a structured knowledge base and careful data preparation in a multilingual fashion, which are of the utmost practical interest for our under-resourced task, that would have otherwise required a substantial additional amount of data. Our model was able to perform and generalise across domains better than a set of strong baselines.
d
e1da9579-ecee-4ebe-9081-a96777759b89
Renewable energy got attention due to the depletion of fossil fuels and global warming. Due to this, the use of DC/DC converters is rapidly increasing in a vast amount of applications such as wind turbines, photovoltaic systems, electric vehicles, energy storage systems, and in such applications, where different voltage levels loads are connected [1]}, [2]}, [3]}. The block diagram of DC MG is expressed in Fig. REF . DC MG mainly includes renewable energy sources (RES) such as solar and wind, energy storage system (ESS), and DC load. Every RES and ESS is connected with the bus through a power electronic interface (PEI). It is therefore necessary to have effective control for the PEIs. The research community extensively proposes different types of linear controllers such as proportional–integral (PI) and proportional–integral–derivative (PID) and is vastly used by the PE industry [4]}, [5]}, [6]}. However, the linear controller has its practical limitations such as tuning of gains, poor disturbance rejection capability, shifting of the operating point of the converter towards instability due to change of the system parameters, and lacking the capacity to handle the non-linearities of the power system. Many nonlinear control techniques such as model predictive control (MPC), sliding mode control (SMC), fuzzy-logic control (FLC) have been proposed to cope with the issues mentioned above and also try to improve the transient behavior. In [7]}, FLC for DC converter is presented for PV-based lighting systems. FLC implementation for DC power converter using microcontroller has been studied in [8]}. FLC basically works on the if-else statement, and its response depends upon predefined rules using if-else logic. FLC does not need any mathematical system model and also has the ability to handle the non-linearity of the system. Voltage regulation of FLC for DC/DC converter is also good under different conditions. However, many studies prove it as an unreliable controller because it lacks formal analysis. So, the amalgam of varying control techniques is found in the literature to balance the disadvantages of FLC [9]}. <FIGURE><FIGURE>
i
8bbe7e21-52ef-4378-bd48-d13f7f8169d9
Slide mode control and model predictive control are developed and vastly studied in literature and have become promising solutions for power electronic converter applications. Slide mode control is based on variable structure control theory. Its basic principle is divided into two stages. The system state trajectory is forcefully taken into the user-defined sliding layer. This phase is known as the reaching phase, then in the second phase, which is known as the sliding phase, state trajectories remain within the layer, defined by the user on the base of application. It has better performance, robust against parametric variations, and possesses magnificent transient response under different loading conditions. Still, chattering phenomena, high switching losses, and complex mathematical modeling are the main barriers to its implementation [1]}, [2]}.
i
180bf802-53a2-4145-895e-bae22fdfce28
MPC is a digital control method, and its basic principle is different from linear control. It uses the discrete-time model of the converter along with its filter to anticipate the behavior for all possible input combinations. One of the inputs having the least (i.e., optimal) value of the predefined cost function (CF) is selected and applied to the coming sampling instant despite drafting a separate loop for each controlled variable and cascading them together as in the case of linear controllers [1]}. CF is basically a square of the Euclidean distance between controlled and reference signals. However, it has a high computational burden, and its performance depends upon the mathematical model of the system and also has variable switching frequency; however, many new studies proposed a constant switching frequency based MPC for different power electronic applications [2]}, [3]}, [4]}.
i
a8efadbc-248e-4db1-91dc-d9c85d41423b
Data-driven or model-free control techniques and especially ANN-based methods are growing in the domain of power converters [1]}. An ANN-based control scheme has been proposed in [2]} to directly control a three-phase inverter with an output LC filter, where a lower THD and a better steady and dynamic performance are achieved. Similarly, authors in [3]} proposed an ANN-based control strategy for a three-phase flying capacitor multi-level inverter (FCMLI). In [4]}, a neural network predictive-based voltage control is proposed for the DC/DC buck converter. The author used PID controller data to train NN. After training, neural network predictive control (NNPC) is used to regulate the voltage. NNPC controller for grid-connected synchronverter is proposed in [5]}.
i
deda7fd1-d225-42a3-a35e-f79e3b5c84b1
They do not require an explicit mathematical model of the system. Their performance is better if they are finely tuned with sufficient data and properly chosen input features [1]}. They can be designed without having expert knowledge.
i
eff56103-ef69-441f-a92d-a0742d411b75
This paper proposes an artificial neural network-based voltage control for a DC/DC step-up (i.e., Boost) converter for DC microgrid applications. Initially, MPC-based voltage control is implemented for the Boost converter to extract the input features data. After the extraction of required data, a different possible combinations of inputs features are chosen. Finally, the voltage reference, inductor current, and capacitor voltage are selected as input features, while the converter switching state is taken as output feature for the proposed ANN in this study, as illustrated in Fig.  REF. Then, these combinations are used to train the ANN. Once the ANN is trained and has good model accuracy, the ANN model is directly used to generate the optimal switching state for the DC converter. Figure REF illustrates the overview of the proposed control strategy: the training phase combines using MPC to anticipate the converter output voltage converter and collection of state variables data under full-state observation. The collected data is used to train the ANN. In the test phase, the trained neural network is employed online to control the converter's output voltage instead of MPC. The simulation results of the proposed control strategy are also compared with the traditional PI Controller.
i
cc1baa12-c8cc-409c-9065-7c9f8c8ed278
The rest of the paper is organized as follows. The mathematical modeling of the DC/DC boost converter and the basic principle of MPC are explained in Section . While the proposed ANN and its training procedure are elaborated in Section . Section shows the simulation results for both ANN and PI controllers. Then, future work is discussed in Section . Finally, Section presents the conclusion. <FIGURE>
i
d4b8fba5-f86c-4d22-bab2-4583035b420e
The trained ANN model is exported into the Simulink model of the DC/DC converter to validate and verify the performance of the proposed control strategy. Extensive MATLAB/Simulink simulation is carried out. The performance of the boost converter with the proposed control scheme is investigated under normal load and step change of load. The simulation parameters of the converter are given in Table REF . <TABLE>
r
c1ff9581-5920-448e-b989-f6c3cf5fb128
Figure REF illustrates the performance of our proposed control strategy, considering normal load conditions. The simulation starts at \(t={0}{}\) , where a resistive load of 20 is connected with the system. The reference voltage is set to 95. Initially, the system takes around 20 to reach the reference value. After a transient period, the output voltage and current wave forms remain stable and do not show distortion. <FIGURE>
r
31e30c9c-9792-4cbf-a133-d12cd5a20e21
Figure REF shows the performance of the proposed controller from full load to no-load condition and vice versa. At \(t={0.4}{}\) , the load is disconnected from the system; i.e., the converter is under no-load condition. It is observed that the voltage remains stable, and the current becomes zero. While at \(t={0.5}{}\) , the load is again connected to the system. The voltage remains stable, while the current is increased to \({4.9}{}\) . However, there is no transient observed in the simulation. After the interval of 0.6, further loads are added into the system to investigate the response of the proposed controller. It is observed from Fig. REF that with increasing the load, the voltage remains stable while keeping track of the reference value with un-noticeable distortion, demonstrating the superior performance of the proposed ANN-based control scheme under different loading and transient conditions. <FIGURE><FIGURE>
r
7f1c71d6-9248-4169-a3b5-0e358ee793e2
Figure REF presents the simulation results of our proposed controller under step change of load. At \(t={0.5}{}\) , the DC load becomes double. We can observe that the voltage remains stable without any distortion, whereas the current increases with the increase of load. However, the current waveform becomes stable with almost no transient time. Figure REF also presents a comparison with the PI controller. Under the transient period, the PI controller shows an overshoot in the voltage and current which may harm the switch of the DC/DC converter; accordingly, a high rating semi-conductor switch is required which increases the converter cost. The ANN-based control scheme has better wave quality and less distortion compared to the PI controller. Moreover, the output current of the PI-based controlled converter is distorted, while the current wave in the case of ANN is constant, stable, and has less loss compared to the PI controller.
r
33acb878-b6ea-4666-b7c6-abf85d1119d6
Within this work, we proposed a feed-forward artificial neural network-based voltage control strategy for the DC/DC step-up converter. Model predictive control is implemented to extract the training data, where the data is used, later on, to train the ANN offline. After training the ANN properly, MPC is removed and the trained ANN successfully regulates the voltage of the DC/DC converter as per reference voltage. The bayesian regularized technique is used to train the ANN and adjust the biases and weights of the ANN. Different types of tests were also performed during simulation, such as step change of load, the shift of load from full load to no load, and vice versa, in order to demonstrate the performance of the proposed controller. It has been observed through simulation results that the overall performance of the proposed control scheme is better than the classical linear controllers. The implementation of the proposed technique would be useful in DC microgrid applications, where the DC boost converters require high accuracy for tuning controller parameters.
d
853c1b26-f28a-473f-9534-16aa9fe7d9d8
Smart contracts are a focal point of modern blockchain environments. Such contracts were firstly popularized by Ethereum [1]}, but soon thereafter other networks developed their own smart contract languages, enabling the implementation of blockchain-based decentralized applications between untrusted parties.
i
06fe6e7e-0c4e-4904-91a8-f4673678dc61
Smart contracts usually operate over user owned assets, thus, vulnerabilities in programs and in the underlying programming languages can lead to considerable losses. The famous attack on the DAO resulted in a theft of approximately 60 million USD worth of Ether [1]}, [2]}, [3]}. Due to recent exploitations of vulnerabilities in smart contracts, blockchain providers turned their attention to the development of robust programming languages, often relying on formal verification, including Liquidityhttp://www.liquidity-lang.org/doc/index.html by Tezos, Plutus by IOHK [4]}, Move by Facebook [5]}, and Rholanghttps://github.com/rchain/rchain/tree/master/rholang-tutorial by RChain. Such languages aim at offering flexible and complex smart contracts while assuring that developers may fully trust contract behaviour. Unfortunately, for Plutus, the last objective has not been completely achieved yet. As we show in the next section, in Plutus, assets can be easily lost forever to the ledger with a simple unintended interaction.
i
f73a96c3-79e8-445f-9855-39b987919650
To counter unplanned interactions with smart contracts endpoints while automating the development of boilerplate code, we propose SmartScribble, a protocol specification language for smart contracts. Its syntax is adapted from the Scribble protocol language [1]} to the smart contract trait and features primitives for sequential composition, choice, recursion, and interrupts. Protocols in SmartScribble specify interactions between participants and the ledger, as well as triggers to interrupt protocol execution. The business logic underlying the contract can be added by the programmer after the automatic generation of the smart contract boilerplate code. The generated code relies on a finite state machine to validate all interactions, precluding unexpected behaviours.
i
52a094e7-8ff0-420f-853c-548eddaf0993
SmartScribble currently targets Plutus, a native smart contract programming language for the Cardano blockchain [1]}, based on the Extended Unspent Transaction Output model [2]}, a solution that expands the limited expressiveness of the Unspent Transaction Output model. In UTxO, transactions consist of a list of inputs and outputs. Outputs correspond to the quantity available to be spent by inputs of subsequent transactions. Extended UTxO expands UTxO's expressiveness without switching to an account-based model, that introduces a notion of shared mutable state, ravelling contract semantics [2]}. Nevertheless, the framework we propose can be integrated with other smart contract languages and blockchain infrastructures expressive enough to support state machines.
i
1940f966-9857-422c-902e-396c53aeb9c8
Several works have been adopting state machines to control the interaction of participants with smart contracts. FSolidM [1]}—the closest proposal to SmartScribble—introduces a model for smart contracts based on finite state machines. FSolidM relies on the explicit construction of finite state machines for contract specification; instead, we automatically generate all state machine code. On a different fashion, the model checker Cubicle [2]} encodes smart contracts and the transactional model of the blockchain as a state machine.
i
8e04424b-c813-48c0-ae34-033c6b8277ce
SmartScribble distinguishes itself from other domain-specific languages—BitML, integrated with the Bitcoin blockchain [1]}, Obsidian [2]}, a typestate-oriented language, and Nomos [3]}, a functional (session-typed) language—by abstracting the interactive behaviour and details of the target programming language through a protocol specification, only relying on the smart contract language to implement the business logic and thus flattening the learning curve.
i
cdae61e5-1789-4a96-85c5-f901d1a2f17f
The next section motivates SmartScribble via an example where assets are lost to the ledger; sec:protocols presents the protocol language and sec:generation contract generation from protocols. sec:evaluation describes some preliminary results of our evaluation of SmartScribble, and sec:conclusion concludes the paper and points to future work. sec:guessgamecode contains the source code for the vulnerable contract we explore in our motivation, sec:logs presents input and logs for playground simulations, sec:approtocols contains the definition of SmartScribble protocols used in sec:evaluation, and sec:plutuscode the source code for the business logic of our running example.
i
647c0a8d-e2e0-4d4f-a111-803b861acc8a
This section compares the amount of lines of code (LOC), those written by the developer and those generated by our compiler. To carry out the comparison, we use the guessing game together with three protocols representative of simple smart contracts. The new protocols are in sec:approtocols.
m
29d0089b-beeb-47a5-89b7-98dbf1e057bf
Ping Pong A simple protocol that alternates between |ping| and |pong| operations, ad eternum. No business logic is required for this protocol. Crowdfunding A crowdfunding where an owner starts a campaign with a goal (in ADA), and contributors donate to the campaign. When the owner decides to close the campaign, all the donations stored in the contract are collected. Auction A protocol where a seller starts an auction over some token, setting the time limit and the maximum number of bids. Buyers bid for the token. When the auction is over, the seller collects the funds of the highest bid and the corresponding bidder gets the token. <TABLE>
m
10c7c981-9b5a-4f43-95f1-ed8bee36130e
table:LOCComp summarizes the analysis. Depending on the protocol, the amount of generated code varies from 150 to 187 lines. In all our examples, the generated code is at least \(10 \times \) larger than the source written in SmartScribble. The business logic varies a lot from contract to contract; nevertheless, it is important to note that it is extremely likely to be a small portion of the complete contract due to the amount of necessary boilerplate that Plutus requires. We see that the ratio between all the code written by the programmer (that is, the protocol and the business logic code) and the Plutus code that would otherwise be manually written is less than 1/4 in all analysed contracts. When we compare LOC for suggested implementationsImplementations available on Plutus' GitHub and SmartScribble, we conclude that with SmartScribble, the code manually written is once again 1/5 or less for every scenario. Even in implementations developed by experts, the boilerplate portion of the contract is significant.
m
952ff768-f3b5-4509-b9c1-48d6210210cb
We present SmartScribble—a protocol language for smart contracts—and a compiler that automatically generates all contract code, except the business logic. The generated code relies on state machines to prevent unexpected interactions with the contract. We claim that SmartScribble improves the reliability of contracts by reducing the likelihood of programmers introducing faults in smart contract code. Our language also flattens the learning curve, allowing developers to focus on the business logic rather than on the boilerplate required to setup a contract, namely in Plutus. Preliminary results point to a 1/4 ratio between the number of lines of code written by the programmer and those in the final contract. This paper constitutes an initial report on using protocol descriptions to generate contract code. Much remains to be done. SmartScribble protocols classify participants under different roles, but we currently do not enforce any form of association of participants to roles. We plan to look into different forms of enforcing the association. Business logic is currently manually written in the contract language (Plutus) and added to the code generated from the protocol. We plan to look into ways of adding more business logic to protocols, thus minimising the Plutus code that must be hand written. Some features of SmartScribble are strongly linked with Plutus. The trigger generation is one of those features: it depends on Plutus libraries for the effect. Nevertheless, we believe that SmartScribble can be adapted to target other languages with minimal changes to the syntax and semantics. Generating Solidity code might be an interesting option for the future, as it also supports state machines. Lastly, evaluation needs to be elaborated. In ongoing work, we are comparing the usage of computational resources between contracts implemented with SmartScribble and the corresponding suggested implementations. Plutus code for vulnerable guessing game language=Haskell,tab=2,showstringspaces=false,breaklines=true]GuessingGame.hs Plutus playground simulation for the guessing game scenario <FIGURE><FIGURE><FIGURE> SmartScribble protocols for the evaluation section The ping-pong protocol protocol PingPongRec (role Client) {   init() from Client;   rec Loop {     ping() from Client;     pong() from Client;     Loop;   } } The crowdfunding protocol protocol Crowdfunding (role Contributor, role Owner){ init (Value) from Owner; rec Loop { choice at Owner{ continue : { contribute (Value) from Contributor; Loop; } closeCrowdfund : {} } } } The auction protocol protocol Auction (role Seller, role Buyer) {   field PubKeyHash, Value;   beginAuction (Token, Value) from Seller {     slot trigger (slot == 10, endAuction);   };   do {     rec Loop {       bid (Value) from Buyer;       Loop;     }   }   interrupt {     endAuction () from Contract;   } } Plutus code for business logic Guessing game logic module (SmartScribble) language=Haskell,tab=2,showstringspaces=false,breaklines=true]GuessingGameLogic.hs
d
3a6460a7-866d-49e4-b2d2-a2143ae97a98
Research has shown that despite the recent slowdown in electrical load growth in the North American energy system, the occurrence of large-scale blackouts has not been reduced [1]}. One contributor to blackouts is the presence of hidden failures in the electrical grid that result in a single fault turning into a multi-element contingency. As described by Elizondo et al. [2]}, a hidden failure is defined as a relay that is misconfigured or faulty such that it will cause the inappropriate removal of system assets during an event (e.g., an out-of-zone fault). It can be stated that hidden failures in protective relaying present a risk of cascading failures to the entire grid.
i
bc749d11-648f-4750-b523-f4207903ffb4
Hidden failures are considered to occur when protection activates on account of a misconfiguration or internal failure, resulting in an inappropriate removal of system assets during an event (e.g., an out-of-zone fault)  [1]}, [2]}. For the most part, hidden failures are synonymous with misoperations and are a consequence of biasing a protection system towards dependability (i.e., always clearing a fault on the protected element) versus security (i.e., never misoperating when a fault has not occurred on the protected element) [3]}. However, the failure of local protection to operate for an in-zone fault, triggering remote backup protection to remove additional system assets, can also be considered a hidden failure [4]}. Of particular concern are hidden failures that cause the loss of a set of critical assets, resulting in a severe loss of load or a cascading failure event, which in turn could lead to prolonged outages and expensive damage.
i
8c3cc922-bdbf-456a-96a3-130b6fa36bbc
While the main cause of power outages remains natural disasters (e.g., extreme weather, forest fires, earthquakes) [1]}, it has been demonstrated that hidden failures present a noticeable impact to the reliability of the United States electrical grid. A number of blackouts are attributed to hidden failures, including the 1977 New York blackout and the 1995 southern Idaho event (in which a line incorrectly tripped on account of a fault on a parallel line, resulting in a subsequent overload trip of a third parallel line, requiring the shed of several GW loads on account of an underfrequency event) [2]}. It is important to mention that hidden failures (including those that cause cascading failures) are responsible for an estimated 10% rate of misoperations across protective device operations [3]} and 70% of \(n-2\) contingencies are caused by relay misoperations [4]}. Additionally, while the electrical grid was designed to be robust to \(n-1\) contingencies, hidden failures can cause multiple-element contingencies for which the risk of load loss or cascading events is much higher [5]}.
i
5179c166-bdaf-4202-84ef-4576bcbbce74
The contributions of this paper include the following: Section  presents a taxonomy of the causes of hidden failures, Section  describes methods for identifying critical locations where hidden failures could pose a risk of cascading failure, and Section  describes methods for mitigation of hidden failures. Although methods in the literature exist for mitigating the impacts of cascading failures – e.g., remedial action schemes [1]} or the “Reciprocal Altruism” method [2]} – this paper specifically focuses on methods to reduce the likelihood of protective misoperations caused by hidden failures.
i
6d47f92d-6a5c-4c20-99be-f0362b192e06
Outages caused by hidden failures are rare compared to those caused by natural disasters (e.g., extreme weather, forest fires, earthquakes). However, research into cascading failures in electrical grids has demonstrated that there are relatively small collections of critical assets that can cause large outages via cascading failures or load shedding from remedial action schemes. The set of assets disabled by a hidden failure can contain one of these collections of critical assets resulting in a large outage as has occurred historically. Given that the number of critical assets in an electrical grid is relatively low, investment in protection system upgrades to reduce the likelihood of hidden failures may likely be cost-effective compared with investments in storm-hardening transmission network infrastructure.
d
98e91469-88d4-4b12-a2d9-20023bc0ac87
The expense in reducing the likelihood of hidden failures may not be high, as several “hidden” failure modes can in fact be detected directly with monitoring, especially those associated with pilot protection. Other failure modes (e.g., shorted timers) are unlikely to occur for integrated digital relays, so the practice of electromechanical relays for backup protection may be inadvisable for critical locations where security is important. For some hidden failures that cannot be detected directly, state-based centralized protection offers a potential solution, especially for applications with nonlinear elements or highly varying operating conditions (e.g., transformer protection or tapped line protection). Other design philosophies can help mitigate the likelihood of misoperations: this includes the use of identical systems for redundancy and the preference for local over remote backup protection.
d
7a678c25-783c-4f9b-849c-be0142a0a16e
T lymphocytes play a vital role in adaptive immunity [1]}, [2]}. The recognition and binding to specific antigens presented on the surface of cells is prerequisite to elicit cytotoxic T cells  [3]}, [4]}. T cells rely on the T cell receptors (TCRs) to recognize antigenic peptides presented by the major histocompatibility complex (MHC) located on the surface of antigen-presenting cells. Human TCR repertoire diversity reach \(10^{16}\) to \(10^{18}\) so that T cells can potentially recognize a huge number of antigens. Previous studies have shown that TCR-antigen binding specificity [5]}, [6]}, [7]} is mainly determined by the heterodimer composed of two \(\alpha \) and \(\beta \) peptide chains [8]}. This variability in specificity stems mostly from plasticity of three complementarity-determining region (CDR) loops (CDR1-3) of both TCR \(\alpha \) - and \(\beta \) -chains, among which CDR3 is responsible for binding to antigenic peptides [1]}, [2]}. So, CDR3 sequence is highly diverse [11]} and the major determinant of antigen binding specificity. Accurate and quantitative estimation of CDR3 diversity and clonal expansion directly reflect the immune response status of T cells to specific antigen stimulation [12]}. TCR-antigen binding specificity is a key factor for the evaluation of therapeutic effect in immunotherapy, such as PD-L1/PD-1 inhibitors, adoptive T-cell immunotherapy and tumor vaccine design. However, the enormous diversity of TCR repertoire makes it a forbidding difficulty for wet-lab experiments to screen out TCR-epitopes binding specificity [13]}. There is a pressing need for development of prediction method to predict TCR binding specificity of antigen, which would greatly complement wet-lab experiments.
i
94747c0c-d74b-4338-b0ed-ea1343eb5125
Several databases with large collection of experimentally validated TCR-peptide interactions, such as VDJdb [1]}, IEDB [2]}, and McPAS [3]}, greatly facilitated the development of computational prediction models. A variety of deep learning-based methods have been developed to predict TCR-antigen binding specificity and acquired encouraging performance [4]}, [5]}. For example, Springer et al. [6]} developed an ERGO model based on natural language processing (NLP) model, using a dictionary of TCR-peptide interactions to predict the binding of TCR and antigenic peptide; Jurtz et al. [7]} developed the NetTCR model based on convolutional neural networks to identify TCRs potentially bound by homologous peptides, making advantage of a large number of non-binding TCR sequences. Lu et al. [8]} established a transfer learning-based model pMTnet to predict binding of TCR to neoantigen presented by MHC I complex. Jun et al.  [9]} proposed a multi-instance learning MIL framework, BERTMHC, to predict antigen binding and presentation by MHC II proteins; Gielis et al.  [10]}developed the web tool TCRex to predict interactions between TCR and epitopes stemmed from multiple types of cancer or viral. These methods have demonstrated the effectiveness of computational prediction of TCR-antigen binding specificity.
i
ac9f4609-043f-4992-91e3-300e03dd237e
Recently, self-supervised representation learning has achieved substantial progress in natural language processing [1]}. The milestone BERT model [2]}, a pre-trained deep bidirectional language representation model based on Transformer [3]}, [2]}, achieved excellent performance on several downstream tasks. Self-supervised contrastive learning [5]} has caught much attention because of its better generalizability in multiple fields. Contrastive learning generated two different views [6]} from the original sample, and trained a deep representation network to minimize the contrastive loss, which encouraged to output similar representations for the augmentations (views) of the same image but dissimilar representations for those generated from other images. Quite a few contrastive learning methods have been proposed, such as SimCLR [5]}, MoCo [8]}, BYOL [9]}, SimSiam [10]}. They run pre-training on large-scale unlabeled dataset, transferred to specific downstream tasks and achieved superior performance and robustness [8]}, [12]}, [13]}. For example, Wang et al. [14]} overcame the drawback of low-quality Position-Specific Scoring Matrix (PSSM) matrix due to poor sequence homology, by running contrastive learning-based feature extraction. Their model achieved great performance improvements on multiple extremely low-quality cases.
i
88c701e6-4460-4883-8a09-02055b6e3806
The success of contrastive learning often relies on the quality of contrastive views. In computer vision, the most commonly used method to generate contrastive views is image augmentation, such as image flipping, random cropping and random Gaussian blur. In graph neural networks, new views were often generated by removal of some nodes, edges or subgraphs. However, all these methods for contrastive view generation pose random perturbation to the original samples. They do not consider the importance of different components of the samples, which may lead to the failure of capturing important information and poor robustness. The feature extraction process spent on non-critical components also results in low computational efficiency.
i
d3e3b872-fe88-4de6-9f7b-026019eb70c7
In this paper, we propose an attention-aware contrastive learning model, ATMTCR, for predicting the binding specificity between TCR and pMHC complex. For each TCR sequence, a proportion of residues were masked under the guidance of attention weights, thereby the contrastive view was generated. In particular, we explored and compared different masking strategies, including masking residues with max- or min- attention weights or roulette probability. We performed exhaustive experiments and demonstrated that contrastive learning significantly improves the predictive performance of TCR-antigen binding. Performance comparison experiments were conducted on two independent TCR-antigen binding datasets, and ATMTCR achieved better performance than other existing algorithms. Moreover, we observed the attention weights help to reveal important amino acids and their positional preferences through, which indicated the interpretability of our proposed model.
i
dac4688d-ab66-4e37-a7a3-dd834c8bdb56
To the best of our knowledge, we were the first to apply contrastive learning for pre-training on large-scale TCR sequences. The obtained TCR sequence representations greatly improved downstream multiple tasks. In addition, our proposed attention masking generates different views by minimizing the contrastive loss of the two views, so that the model focuses on key positions and specific amino acids in the sequence, and successfully extracts high-order semantic information from the sequence.
i
d34aca88-24e1-401f-86ce-dccc05de8517
Understanding the binding of TCR and antigenic peptides plays a key role in the estimation of the efficacy of immune checkpoint inhibitors and development of immunotherapies. However, due to the high polymorphism of TCR sequences, it is difficult to capture the biological mechanism of antigenic peptides eliciting T cells. On the other hand, the incredible diversity of TCR repertoire make it impractical to screen TCR-antigen bindings via wet-lab experiments.
d
a2d13b42-7275-49c3-af35-ed127b32d62b
Model-based prediction of TCR-antigen binding also run into difficulty due to the paucity of labeled data. Self-supervised learning showed promising progress in learning informative representations by designing some pretext tasks. We adopted this new technique to perform representation learning of TCR sequences, and meanwhile tied to increase the interpretability of the model by using self-attention mechanism.
d
bcd02181-c7a6-4895-8235-296a228a1b28
As a result, we observed interesting appearance and distribution patterns of some amino acids that potentially determine TCR-epitope binding specificity. We have drawn the conclusion that specific amino acids localized in key region of the CDR loops would bind strongly to antigenic peptides and MHC molecules, thereby eliciting immune response. Lin X et al. [1]} have delineated the 3Dl crystal structure of TCR-pMHC complex, and showed some regions were more attractive to antigenic peptide. From this perspective, we hope to extend the model in our future work to narrow the region where TCR binds to antigenic peptides for triggering immune responses, thereby reducing the experimental cost.
d
ac44a1d3-8c52-402d-aeb6-2d9759374f0c
In this study, we performed self-supervised representation learning on a large number of TCR sequences, and fine-tuned the model on experimentally validated TCR-epitope data. Our experimental verified our model achieved promising results. By pretraining on a variety of TCR sequences, the model allows us to better understand the determinant factors of TCRs in binding to epitopes, as well as the importance of the HLA type for the presentation of epitopes. Given more data, we believe our model is possible to predict whether TCRs recognize previously unseen epitopes.
d
4103cc3b-3fd6-4bce-ad2c-818c3fca4f64
Large-scale language technologies are increasingly used to enable various forms of linguistic communication in contexts ranging from biomedical research to education to machine translation [1]}, [2]}, [3]}. A particular class of these technologies, conversational agents, primarily engage in linguistic communication with humans by outputting natural language text in response to prompts and queries.We use the term `conversational agents' as suggested by Perez-Marin and Pascual-Nieto [4]}. These technologies are also known as `dialogue systems' [5]}. Central to their performance is the development of large language models, such as GPT-3 or BERT, which analyse text data and employ statistical techniques to determine the probability distribution of a sequence of text.For GPT-3, see [2]}; for BERT, see [7]}; for Turing-NLG-A-17, see [8]}; for CLIP, see [9]}; for Gopher, see [10]}. These models are trained on a vast corpus of text-based materials, ranging from Wikipedia articles to online repositories of computer code. They can then be adapted to perform a range of different conversational tasks.
i
9636833d-2e92-42bb-8fa0-c47fe3874f00
Conversational agents have been shown to perform well on a variety of computational metrics, supporting the emergence of new kinds of capability and opportunity [1]}, [2]}.For example, the MMLU (Multi-task Language Understanding) and MATH datasets each consist of a set of problems and solutions that are central to human knowledge.These datasets are used to evaluate whether language models can correctly generate solutions to these problems. However, early instances of these models also present a number of risks and possible failure modes, including the production of false, offensive, or irrelevant information that could lead to a range of harms [3]}, [4]}, [5]}. A key social and ethical issue that arises in this context, therefore, concerns the alignment of conversational agents with appropriate norms and values.For an in-depth examination of value alignment, see Gabriel [6]}. Which standards, if any, should conversational agents be aligned with, and how can this be accomplished?
i
4202a7ce-94f3-4c3f-92f6-4cd704265cc5
To date, efforts to create aligned conversational agents have centred on the identification and mitigation of harms, such as the proliferation of inappropriate stereotypes or of hateful speech [1]}, [2]}, [3]}. These responses focus on providing solutions to particular problems in the hope that their reduction or elimination will lead to the creation of good or beneficial conversational agents that no longer cause harm. Yet, while this `bottom-up' approach is useful in tackling specific harms, we cannot assume that the piecemeal elimination of unwanted outcomes will necessarily lead to the creation of language technologies that are substantively beneficial.One reason for this stems from the fact that the cultivation of virtues is not necessarily equivalent to the elimination of errors. Certain virtues may be supererogatory and hence desirable but not morally required. In these cases, the absence of virtue leads not to harm but to a failure to realise better states of affairs. Taken on its own, the bottom-up approach risks `patching' certain problems but leaving other questions about the design of conversational agents – such as the meaning of `good speech' – largely untouched.
i
bc2a69bd-b353-4dce-a5bb-90ef23721b46
For example, there is widespread agreement that language models output false and low-quality information [1]}. However, this observation leads quite naturally to the question of what it means for an utterance to be truthful. Does the same notion of truth apply across a wide range of conversational contexts, or might standards of truthfulness vary according to the subject under consideration and to relevant conversational norms? Equally, there is widespread concern that the output of large-scale language models is biased [2]}, [3]}. Yet this raises the question of what it means for language models to be unbiased. Moreover, when is the goal of producing unbiased language appropriate? And what conception of bias, among the plurality of options, ought to serve as the focal point for corrective action?See, for example, Mehrabi et al. [4]}, Mitchel et al. [5]}, and Kasirzadeh [6]}.
i
7fa1013b-8124-46fc-aa27-713378079952
To address these issues properly, we need to draw upon a second complementary perspective. The `top-down' approach to conversational agent alignment focuses on identifying general principles or ideals that can successfully guide linguistic communication.See Allen et al. [1]} for an introduction to a top-down approach. This approach seeks to specify more precisely what ideal linguistic communication is, across a range of contexts, and to realise these properties in the design of conversational agents.There is also a second functional meaning of `good speech' which is defined at a higher-level of abstraction (See Kasirzadeh and Klein [2]} for levels of abstraction in machine learning systems). This meaning also needs to be satisfied by a top-down approach. For example, a conversational agent that is supposed to output an accurate summary of texts, outputs `good speech' if it provides an accurate summary of texts. This paper explores how a top-down approach might be developed and implemented, in order to complement the bottom-up approach discussed already.
i
8f518cad-5584-49ab-9aa1-da053de4e6b4
We start by exploring three types of requirements that plausibly need to be satisfied for successful human-conversational agent communication to take place: these are, syntactic, semantic, and pragmatic criteria (Section 2).In this paper we use the term pragmatics to encompass both the focus on a single utterance as well as discourse more broadly. While syntactic and semantic norms have been widely examined in the present context, the nature and significance of pragmatic norms for successful discourse between humans and conversational agents has received less attention in large language model scholarship.We would like to acknowledge that the pre-neural network literature about pragmatic norms for natural language generation includes careful and thoughtful scholarship on this area. See, for example, Dale and Reiter [1]} and Asher et al. [2]}. We thank Ben Hutchinson for bringing these resources to our attention. Our focus in this paper is post-neural network literature and in particular the domain of large-scale language models. To remedy this situation, we delve deeper into the components of successful linguistic communication and show why, and in what ways, pragmatic concerns and norms are central to the design of aligned conversational agents (Section 3). Language performs many roles and functions in different contexts. Therefore, an account of successful communication also needs to consider whether an utterance is valuable in relation to what end, for which group of people, and in what way. To answer these questions, we examine how an additional set of norms which we call discursive ideals contribute to the success of a conversation in specific domains. In particular, we explore what these discursive ideals look like in the context of scientific discourse, democratic debate, and the process of creative exchange (Section 4). We then look at the practical implications of our discussion for future research on value-aligned conversational agents and consider whether the approach developed here captures all – or even the most important – values underlying the design of successful conversational agents (Section 5). The paper concludes with some final remarks (Section 6).
i
fdc6e140-2c3b-4863-b61e-ce871500f45a
Related work and novelty. Before we go further, we would like to take stock of existing work on the alignment of large-scale language technologies, to highlight the novelty and potential utility of our own contribution to this important area of scholarship, debate, and practice.
i
3d542796-ff4f-436c-a8a6-cc35378e8849
Henderson et al. [1]} explore the kinds of ethical and safety problems that might arise from the use of dialogue agents in conversation with human interlocutors. However, they do not propose a solution to these challenges they identify. Additionally, three recent papers by Bender et al. [2]}, Tamkin et al. [3]}, and Weidinger et al. [4]} explore the many kinds of harm that can be produced by language models, such as misinformation, bias, and inappropriate anthropomorphism. We share these concerns and commend the value of the research. Yet, we also think that it is important to move beyond problem identification and the mitigation of individual risks, and to complement these efforts with further normative reflection about what appropriate or ideal conversation looks like. In this paper, we propose a number of steps that support this task, in order to better align conversational agents with human values and to develop a more holistic account of what aligned non-harmful conversational agents might look like.
i
142899b2-34ee-4b98-8ff9-92d031524301
Drawing upon the AI-safety literature, Kenton et al. [1]} explore the ways in which human designers may misspecify objectives for conversational agents and propose mitigation strategies to avoid the creation of accidentally misaligned conversational agents. While these strategies are necessary, we think that they are not sufficient to ensure the creation of aligned language technologies. In this paper, we provide a fuller account of the goals and standards conversational agents ought to be aligned with in the first place. Evans et al. [2]} propose mitigation strategies to avoid several kinds of falsehood produced by language technologies. In particular, their work demonstrates an important way in which a norm around truthfulness might be operationalised. However, their conception of falsehood does not map equally well onto all contexts and needs to be supplemented by other standards in, for example, democratic or creative settings. Finally, Askell et al. [3]} propose three ideals – helpfulness, honesty, and harmlessness – that can help align conversational agents with human values. Their work points towards several important norms that apply across a range of contexts; however, as with Evans et al. [2]}, these norms do not apply equally to all contexts in which conversational agents might be productively deployed.
i
af2da5e9-e579-4dfa-94cb-a11c5ec6d3bc
In sum, we believe that, by paying attention to the context of a conversation and the norms that structure the conversational domain, it is possible to develop conversational agents that produce beneficial communication with human interlocutors. We demonstrate the need for research into appropriate norms of conversation that govern interaction between conversational agents and human interlocutors, and we provide illustrations of what this might look like. Finally, we would like to acknowledge that the primary focus of our paper is on ideal speech for the English language.The English language is not itself monolithic, containing many varieties, areas of contestation and sets of sociolinguistic relationships. Nonetheless, for the sake of simplicity, and in order to convey our points more clearly, we talk about the English language in this paper. We do not discuss how our arguments can carry over to other languages or different modes of communication such as oral, rather than written, linguistic traditions.There are a variety of non-English language models [1]}. Moreover, multilingual language modelling is an important and budding research area [2]}. We believe it is an important open question – and one for further research – whether and in what way other languages, language varieties, and cultural traditions, may generate different interpretations of the normative ideals that inform speech and communication.
i
c108aa8c-f0ca-44f0-97d3-cdb6fa86731c
This paper addresses the alignment of large-scale conversational agents with human values. Drawing upon philosophy and linguistics we explored components of successful linguistic communication (with a focus on English textual language) and showed why and in what ways pragmatic norms and concerns are central to the design of ideal conversational agents. Building upon these insights, we mapped out a set of discursive ideals governing three different domains, in order to illustrate how pragmatic theory can inform the design of aligned conversational agents. These ideals, in conjunction with Gricean maxims, comprise one way in which the top-down approach to design of more ideal conversational agents can be operationalised. For each discursive domain, our overview of the relevant norms was impressionistic; the interpretation of operationalisation of these norms requires further technical and non-technical investigation. Indeed, as our analysis made clear, the norms embedded in different cooperative practices – whether those of science, civic life, or creative exchange – must themselves be subjected to reflective evaluation and critique [1]}, [2]}. Lastly, we highlighted some practical implications of our proposal with respect to future research on the design of ideal conversational agents and human–language agent conversations. These include the need for a context-sensitive evaluation and fine-tuning of language models, and our hope that, once aligned with relevant values, these models can help nurture more productive forms of conversational exchange. Our focus in this paper has been on the English language and the alignment of conversational agents with a particular set of communicative norms in related contexts. In addition to deeper investigation of the norms proposed herein, a complementary exploration of the norms that structure other languages and linguistic traditions is an important task that needs to be explored elsewhere.
d
03e16429-ecb3-477a-acb0-27bd29baaaf0
Over the past decade, deep neural networks have shown superior performances in various domains, such as visual recognition, natural language processing, robotics, and healthcare. However, recent studies have demonstrated that machine learning models are vulnerable in terms of leaking private data [1]}, [2]}, [3]}. Hence, preventing private data from being recovered by malicious attackers has become an important research direction in deep learning research.
i
257a9097-2f5d-4fea-b7fa-199c12054e1f
Distributed machine learning [1]}, [2]} has emerged as an attractive setting to mitigate privacy leakage without requiring clients to share raw data. In the case of an edge-cloud distributed learning scenario, most layers are commonly offloaded to the cloud, while the edge device computes only a small number of convolutional layers for feature extraction, due to power and resource constraints [3]}. For example, service provider trains and splits a neural network at a “cut layer,” then deploys the rest of the layers to clients [4]}. Clients encode their dataset using those layers, then send the data representations back to cloud server using the rest of layers for inference [5]}, [6]}, [4]}. This gives an untrusted cloud provider or a malicious participant a chance to steal sensitive inference data from the output of “cut layer” on the edge device side, i.e. inverting data from their outputs [8]}, [9]}.
i
a42b3a83-c9bb-4742-aad3-3504fab98a2e
In this paper, we investigate how to design a hard-to-invert data representation function (or hidden data representation function), which is defined as the output of the neural network's intermediate layer. We focus on defending data recovery during inference. The goal is to hide sensitive information and to protect data representations from being used to reconstruct the original data while ensuring that the resulted data representations are still informative enough for decision making. The core question here is how to achieve the goal.
i