Director of Machine Learning Insights [Part 1]
Few seats at the Machine Learning table span both technical skills, problem solving and business acumen like Directors of Machine Learning
Directors of Machine Learning and/or Data Science are often expected to design ML systems, have deep knowledge of mathematics, familiarity with ML frameworks, rich data architecture understanding, experience applying ML to real-world applications, solid communication skills, and often expected to keep on top of industry developments. A tall order!
For these reasons, we’ve tapped into this unique group of ML Directors for a series of articles highlighting their thoughts on current ML insights and industry trends ranging from Healthcare to Finance, eCommerce, SaaS, Research, Media, and more. For example, one Director will note how ML can be used to reduce empty deadheading truck driving (which occurs ~20% of the time) down to just 19% would cut carbon emissions by ~100,000 Americans. Note: This is back of napkin math, done by an ex-rocket Scientist however, so we’ll take it.
In this first installment, you’ll hear from a researcher (who’s using ground penetrating radar to detect buried landmines), an ex-Rocket Scientist, a Dzongkha fluent amateur gamer (Kuzu = Hello!), an ex-van living Scientist, a high-performance Data Science team coach who’s still very hands-on, a data practitioner who values relationships, family, dogs, and pizza. —All of whom are currently Directors of Machine Learning with rich field insights.
🚀 Let’s meet some top Machine Learning Directors and hear what they have to say about Machine Learning’s impact on their prospective industries:
Archi Mitra - Director of Machine Learning at Buzzfeed
Background: Bringing balance to the promise of ML for business. People over Process. Strategy over Hope. AI Ethics over AI Profits. Brown New Yorker.
Buzzfeed: An American Internet media, news and entertainment company with a focus on digital media.
Privacy first personalization for customers: Every user is unique and while their long-term interests are stable, their short-term interests are stochastic. They expect their relationship with the Media to reflect this. The combination of advancement in hardware acceleration and Deep Learning for recommendations has unlocked the ability to start deciphering this nuance and serve users with the right content at the right time at the right touchpoint.
Assistive tools for makers: Makers are the limited assets in media and preserving their creative bandwidth by ML driven human-in-the-loop assistive tools have seen an outsized impact. Something as simple as automatically suggesting an appropriate title, image, video, and/or product that can go along with the content they are creating unlocks a collaborative machine-human flywheel.
Tightened testing: In a capital intensive media venture, there is a need to shorten the time between collecting information on what resonates with users and immediately acting on it. With a wide variety of Bayesian techniques and advancements in reinforcement learning, we have been able to drastically reduce not only the time but the cost associated with it.
Privacy, editorial voice, and equitable coverage: Media is a key pillar in the democratic world now more than ever. ML needs to respect that and operate within constraints that are not strictly considered table stakes in any other domain or industry. Finding a balance between editorially curated content & programming vs ML driven recommendations continues to be a challenge. Another unique challenge to BuzzFeed is we believe that the internet should be free which means we don't track our users like others can.
Ignoring “the makers” of media: Media is prevalent because it houses a voice that has a deep influence on people. The editors, content creators, writers & makers are the larynx of that voice and the business and building ML that enables them, extends their impact and works in harmony with them is the key ingredient to success.
Hopefully, small data-driven general-purpose multi-modal multi-task real-time ML systems that create step-function improvements in drug discovery, high precision surgery, climate control systems & immersive metaverse experiences. Realistically, more accessible, low-effort meta-learning techniques for highly accurate text and image generation.
Li Tan - Director of Machine Learning & AI at Johnson & Johnson
Background: Li is an AI/ML veteran with 15+ years of experience leading high-profile Data Science teams within industry leaders like Johnson & Johnson, Microsoft, and Amazon. Li continues to be curious, is always learning and enjoys hands-on programming.
Fun Fact: Li continues to be curious, is always learning, and enjoys hands-on programming.
Johnson & Johnson: A Multinational corporation that develops medical devices, pharmaceuticals, and consumer packaged goods.
AI/ML applications have exploded in the pharmaceuticals space the past few years and are making many long-term positive impacts. Pharmaceuticals and healthcare have many use cases that can leverage AI/ML.
Applications range from research, and real-world evidence, to smart manufacturing and quality assurance. The technologies used are also very broad: NLP/NLU, CV, AIIoT, Reinforcement Learning, etc. even things like AlphaFold.
The biggest ML challenge within pharma and healthcare is how to ensure equality and diversity in AI applications. For example, how to make sure the training set has good representations of all ethnic groups. Due to the nature of healthcare and pharma, this problem can have a much greater impact compared to applications in some other fields.
Wouldn’t say this is necessarily a mistake, but I see many people gravitate toward extreme perspectives when it comes to AI applications in healthcare; either too conservative or too aggressive.
Some people are resistant due to high regulatory requirements. We had to qualify many of our AI applications with strict GxP validation. It may require a fair amount of work, but we believe the effort is worthwhile. On the opposite end of the spectrum, there are many people who think AI/Deep Learning models can outperform humans in many applications and run completely autonomously.
As practitioners, we know that currently, neither is true.
ML models can be incredibly valuable but still make mistakes. So I recommend a more progressive approach. The key is to have a framework that can leverage the power of AI while having goalkeepers in place. FDA has taken actions to regulate how AI/ML should be used in software as a medical device and I believe that’s a positive step forward for our industry.
The intersections between AI/ML and other hard sciences and technologies. I’m excited to see what’s to come.
Alina Zare - Director of the Machine Learning & Sensing Laboratory at the University of Florida
Background: Alina Zare teaches and conducts research in the area of machine learning and artificial intelligence as a Professor in the Electrical and Computer Engineering Department at the University of Florida and Director of the Machine Learning and Sensing Lab. Dr. Zare’s research has focused primarily on developing new machine learning algorithms to automatically understand and process data and imagery.
Her research work has included automated plant root phenotyping, sub-pixel hyperspectral image analysis, target detection, and underwater scene understanding using synthetic aperture sonar, LIDAR data analysis, Ground Penetrating Radar analysis, and buried landmine and explosive hazard detection.
Fun Fact: Alina is a rower. She joined the crew team in high school, rowed throughout college and grad school, was head coach of the University of Missouri team while she was an assistant professor, and then rowed as a masters rower when she joined the faculty at UF.
Machine Learning & Sensing Laboratory: A University of Florida laboratory that develops machine learning methods for autonomously analyzing and understanding sensor data.
ML has made a positive impact in a number of ways from helping to automate tedious and/or slow tasks or providing new ways to examine and look at various questions. One example from my work in ML for plant science is that we have developed ML approaches to automate plant root segmentation and characterization in imagery. This task was previously a bottleneck for plant scientists looking at root imagery. By automating this step through ML we can conduct these analyses at a much higher throughput and begin to use this data to investigate plant biology research questions at scale.
There are many challenges. One example is when using ML for Science research, we have to think carefully through the data collection and curation protocols. In some cases, the protocols we used for non-ML analysis are not appropriate or effective. The quality of the data and how representative it is of what we expect to see in the application can make a huge impact on the performance, reliability, and trustworthiness of an ML-based system.
Related to the question above, one common mistake is misinterpreting results or performance to be a function of just the ML system and not also considering the data collection, curation, calibration, and normalization protocols.
There are a lot of really exciting directions. A lot of my research currently is in spaces where we have a huge amount of prior knowledge and empirically derived models. For example, I have ongoing work using ML for forest ecology research. The forestry community has a rich body of prior knowledge and current purely data-driven ML systems are not leveraging. I think hybrid methods that seamlessly blend prior knowledge with ML approaches will be an interesting and exciting path forward. An example may be understanding how likely two species are to co-occur in an area. Or what species distribution we could expect given certain environmental conditions. These could potentially be used w/ data-driven methods to make predictions in changing conditions.
Nathan Cahill Ph.D. - Director of Machine Learning at Xpress Technologies
Background: Nathan is a passionate machine learning leader with 7 years of experience in research and development, and three years experience creating business value by shipping ML models to prod. He specializes in finding and strategically prioritizing the business' biggest pain points: unlocking the power of data earlier on in the growth curve.
Fun Fact: Before getting into transportation and logistics I was engineering rockets at Northrop Grumman. #RocketScience
Xpress Technologies: A digital freight matching technology to connect Shippers, Brokers and Carriers to bring efficiency and automation to the Transportation Industry.
The transportation industry is incredibly fragmented. The top players in the game have less than 1% market share. As a result, there exist inefficiencies that can be solved by digital solutions.
For example, when you see a semi-truck driving on the road, there is currently a 20% chance that the truck is driving with nothing in the back. Yes, 20% of the miles a tractor-trailer drives are from the last drop off of their previous load to their next pickup. The chances are that there is another truck driving empty (or "deadheading") in the other direction.
With machine learning and optimization this deadhead percent can be reduced significantly, and just taking that number from 20% to 19% percent would cut the equivalent carbon emissions of 100,000 Americans.
Note: the carbon emissions of 100k Americans were my own back of the napkin math.
The big challenge within logistics is due to the fact that the industry is so fragmented: there is no shared pool of data that would allow technology solutions to "see" the big picture. For example a large fraction of brokerage loads, maybe a majority, costs are negotiated on a load by load basis making them highly volatile. This makes pricing a very difficult problem to solve. If the industry became more transparent and shared data more freely, so much more would become possible.
I think that the most common mistake I see is people doing ML and Data Science in a vacuum.
Most ML applications within logistics will significantly change the dynamics of the problem if they are being used so it's important to develop models iteratively with the business and make sure that performance in reality matches what you expect in training.
An example of this is in pricing where if you underprice a lane slightly, your prices may be too competitive which will create an influx of freight on that lane. This, in turn, may cause costs to go up as the brokers struggle to find capacity for those loads, exacerbating the issue.
I think the thing that excites me most about ML is the opportunity to make people better at their jobs.
As ML begins to be ubiquitous in business, it will be able to help speed up decisions and automate redundant work. This will accelerate the pace of innovation and create immense economic value. I can’t wait to see what problems we solve in the next 10 years aided by data science and ML!
Nicolas Bertagnolli - Director of Machine Learning at BEN
Background: Nic is a scientist and engineer working to improve human communication through machine learning. He’s spent the last decade applying ML/NLP to solve data problems in the medical space from uncovering novel patterns in cancer genomes to leveraging billions of clinical notes to reduce costs and improve outcomes.
Fun Fact: Nic lived in a van and traveled around the western United States for three years before starting work at BEN.
BEN: An entertainment AI company that places brands inside influencer, streaming, TV, and film content to connect brands with audiences in a way that advertisements cannot.
In so many ways! It’s completely changing the landscape. Marketing is a field steeped in tradition based on gut feelings. In the past 20 years, there has been a move to more and more statistically informed marketing decisions but many brands are still relying on the gut instincts of their marketing departments. ML is revolutionizing this. With the ability to analyze data about which advertisements perform well we can make really informed decisions about how and who we market to.
At BEN, ML has really helped us take the guesswork out of a lot of the process when dealing with influencer marketing. Data helps shine a light through the fog of bias and subjectivity so that we can make informed decisions.
That’s just the obvious stuff! ML is also making it possible to make safer marketing decisions for brands. For example, it’s illegal to advertise alcohol to people under the age of 21. Using machine learning we can identify influencers whose audiences are mainly above 21. This scales our ability to help alcohol brands, and also brands who are worried about their image being associated with alcohol.
As with most things in Machine Learning the problems often aren’t really with the models themselves. With tools like Hugging Face, torch hub, etc. so many great and flexible models are available to work with.
The real challenges have to do with collecting, cleaning, and managing the data. If we want to talk about the hard ML-y bits of the job, some of it comes down to the fact that there is a lot of noise in what people view and enjoy. Understanding things like virality are really really hard.
Understanding what makes a creator/influencer successful over time is really hard. There is a lot of weird preference information buried in some pretty noisy difficult-to-acquire data. These problems come down to having really solid communication between data, ML, and business teams, and building models which augment and collaborate with humans instead of fully automating away their roles.
I don’t think this is exclusive to marketing but prioritizing machine learning and data science over good infrastructure is a big problem I see often. Organizations hear about ML and want to get a piece of the pie so they hire some data scientists only to find out that they don’t have any infrastructure to service their new fancy pants models. A ton of the value of ML is in the infrastructure around the models and if you’ve got trained models but no infrastructure you’re hosed.
One of the really nice things about BEN is we invested heavily in our data infrastructure and built the horse before the cart. Now Data Scientists can build models that get served to our end users quickly instead of having to figure out every step of that pipeline themselves. Invest in data engineering before hiring lots of ML folks.
There is so much exciting stuff going on. I think the pace and democratization of the field is perhaps what I find most exciting. I remember almost 10 years ago writing my first seq2seq model for language translation. It was hundreds of lines of code, took forever to train and was pretty challenging. Now you can basically build a system to translate any language to any other language in under 100 lines of python code. It’s insane! This trend is most likely to continue and as the ML infrastructure gets better and better it will be easier and easier for people without deep domain expertise to deploy and serve models to other people.
Much like in the beginning of the internet, software developers were few and far between and you needed a skilled team to set up a website. Then things like Django, Rails, etc. came out making website building easy but serving it was hard. We’re kind of at this place where building the models is easy but serving them reliably, monitoring them reliably, etc. is still challenging. I think in the next few years the barrier to entry is going to come WAY down here and basically, any high schooler could deploy a deep transformer to some cloud infrastructure and start serving useful results to the general population. This is really exciting because it means we’ll start to see more and more tangible innovation, much like the explosion of online services. So many cool things!
Eric Golinko - Director of Machine Learning at E Source
Background: Experienced data practitioner and team builder. I’ve worked in many industries across companies of different sizes. I’m a problem solver, by training a mathematician and computer scientist. But, above all, I value relationships, family, dogs, travel and pizza.
Fun Fact: Eric adores nachos!
E Source: Provides independant market intelligence, consulting, and predictive data science to utilities, major energy users, and other key players in the retail energy marketplace.
Access to business insight. Provided a pre-requisite is great data. Utilities have many data relationships within their data portfolio from customers to devices, more specifically, this speaks to monthly billing amounts and enrollment in energy savings programs. Data like that could be stored in a relational database, whereas device or asset data we can think of as the pieces of machinery that make our grid. Bridging those types of data is non-trivial.
In addition, third-party data spatial/gis and weather are extremely important. Through the lens of machine learning, we are able to find and explore features and outcomes that have a real impact.
There is a demystification that needs to happen. What machine learning can do and where it needs to be monitored or could fall short. The utility industry has established ways of operating, machine learning can be perceived as a disruptor. Because of this, departments can be slow to adopt any new technology or paradigm. However, if the practitioner is able to prove results, then results create traction and a larger appetite to adopt. Additional challenges are on-premise data and access to the cloud and infrastructure. It’s a gradual process and has a learning curve that requires patience.
Not unique to utilizes, but moving too fast and neglecting good data quality and simple quality checks. Aside from this machine learning is practiced among many groups in some direct or indirect way. A challenge is integrating best development practices across teams. This also means model tracking and being able to persist experiments and continuous discovery.
I’ve been doing this for over a decade, and I somehow still feel like a novice. I feel fortunate to have been part of teams where I’d be lucky to be called the average member. My feeling is that the next ten years and beyond will be more focused on data engineering to see even a larger number of use cases covered by machine learning.
🤗 Thank you for joining us in this first installment of ML Director Insights. Stay tuned for more insights from ML Directors in SaaS, Finance, and e-Commerce.
Big thanks to Eric Golinko, Nicolas Bertagnolli, Nathan Cahill, Alina Zare, Li Tan, and Archi Mitra for their brilliant insights and participation in this piece. We look forward to watching each of your continued successes and will be cheering you on each step of the way. 🎉
Lastly, if you or your team are interested in accelerating your ML roadmap with Hugging Face Experts please visit hf.co/support to learn more.