On October 18, 2017, Object Computing Principal Software Engineer Mike Martinez joined fellow software engineers, data scientists, and machine learning practitioners from across the country in an engaging exploration of artificial intelligence (AI) at the inaugural StampedeCon AI Summit in St. Louis, Missouri.
As an engineer and a software developer, Mike has been using high-volume, high-velocity data to solve complex problems for over three decades. He attended the StampedeCon AI Summit to network with other professionals and get a glimpse of the technological advances on the horizon.
Eighteen machine learning and AI experts presented talks on a wide variety of topics at the Summit. Mike attended a number of these presentations and shared what he learned. Here are the talks and speakers he found most interesting and some of the key ideas he brought back with him.
Bhaskar Dutta: Why Should We Trust You?
In “Why Should We Trust You? Interpretability of Deep Neural Networks,” Bhaskar Dutta, a Monsanto data scientist and engineer who works in computer vision and machine learning, talked about the mechanisms used to interpret the models that result from training deep neural networks.
Dutta said this topic is particularly relevant today as the European Union (EU) renews its focus on transparency in government and business environments. With the advent of opaque (black box) deep learning network models, compliance with legal transparency requirements can be a challenge.
Legal concerns aside, Dutta said the ability to explain and interpret models also increases trust in the model results. This is important because model results are often counterintuitive.
One technique that can be used to interpret results is called LIME (Local Interpretable Model-Agnostic Explanations). LIME relies on a practice of splitting the model or feature space into smaller sections, each of which can be more easily evaluated and interpreted. For example, if a decision is made to classify a credit risk as good or bad, interpreting the model’s entire feature set is unnecessary; only the local boundary in the decision space needs to be examined to explain the result.
Igor Zwir: Bringing the Whole Elephant Into View
Igor Zwir, PhD, associate professor for the Department of Computer Science and Artificial Intelligence at the University of Granada, Spain, and founder of D&C Intelligent Systems, presented “Bringing the Whole Elephant Into View: Can Cognitive Systems Bring Real Solutions to Complex Problems?”
Zwir’s presentation focused on the fact that there is currently no commonly agreed-upon definition of a cognitive system. He asserted that before cognitive systems can become a viable tool for solving challenging problems in medicine, economics, and other fields, it is necessary to develop an understanding of the ways in which all the various facets of such a system work together to create a whole.
Zwir, who is also an Assistant Professor at Washington University School of Medicine’s Department of Psychiatry, was a senior investigator on a team that discovered, through experiments in machine learning, that there are at least eight types of schizophrenia. The conclusions reached by Zwir’s team were based upon a realization that loss of detail due to aggregation and averaging, as well as inclusion of information from additional sources (such as the psychological and physical domains), allowed for better classifications.
Zwir presented these findings as a cautionary tale for those who rely upon pooling data to calculate results instead of utilizing all of the available information. He stated that because computing and storage resources are now available to handle the scale for many of these models, it is important that models are trained to use all the data available, rather than aggregating or otherwise pooling it.
Harrison Knoll: Geospatial Artificial Intelligence
In a talk on “Geospatial Artificial Intelligence,” Harrison Knoll, experimental physicist and CEO of Aerial Insights, a startup that analyzes drone imagery, discussed feature extraction and recognition processes.
Knoll said most imaging processes are currently performed by models trained on the CIFAR dataset or other similar sets. He said if a person wants to distinguish between dogs and cats, these data sets do a great job. But, if one needs to recognize and classify other types of objects via images, machines must be trained with a completely new, labeled data set. This is time consuming and prone to problems that users of recognition software do not typically anticipate.
Knoll discussed a current project in which his company flies drones near electrical transmission towers to inspect insulators. This significantly reduces the time it takes to complete inspections, and it has the added benefit of risk reduction for employees who no longer have to climb the towers and perform the inspections manually.
In this case, before Knoll’s team could implement the solution, they had to acquire and extend imagery to classify insulators and train the drones to determine whether the insulators, or any of their parts, were intact or broken.
S. Joshua Swamidass: Demystifying Deep Learning
The highlight of the StampedeCon AI Summit was a presentation on “Demystifying Deep Learning” by keynote speaker, S. Joshua Swamidass, MD, PhD, who is an Assistant Professor of Laboratory and Genomic Medicine and the faculty lead for the Translational Bioinformatics Institute for Informatics at Washington University.
Swamidass used a metaphor involving a piano to describe the current machine learning landscape.
He said that if everyone had a piano in the living room, such a scenario would be similar to machine learning today. Everyone knows about pianos; everyone knows what they look like and what they’re supposed to do. But not everyone understands what a piano is truly capable of or how to play one. Only a select few learn to play entire classical pieces to perfection.
According to Swamidass, those who become master pianists and learn to flawlessly play classical compositions can be likened to future machine learning experts. They not only understand the tools they’re working with, they also have the skills to use them in extraordinary ways.
Swamidass said that for those interested in improving their understanding of and ability to use machine learning and AI, resources are currently scarce, and that this is unlikely to improve for some time. His recommendation to companies interested in leveraging these technologies is to seek out consultants; this approach allows companies to field the specific capabilities they need.
Swamidass also made a distinction between researchers and practitioners. He said researchers focus their efforts on the theory behind machine learning, AI capabilities, and potential solutions that might be derived from the technology. Practitioners are the ones who use the researcher’s findings for practical purposes.
In other words, it is the machine learning engineers – not researchers – who will develop, scale, and deliver data and machine learning products that help companies benefit from this burgeoning field.
Based upon Swamidass’ distinction, it seems reasonable to conclude that engineers should enter the field as consulting practitioners rather than as researchers if they are interested in delivering the leading-edge solutions the technology will one day support.
Final Takeaways
Mike said that the StampedeCon AI Summit was a great opportunity to network with software engineers and data scientists with an interest in exploring new technological horizons. The presentations were led by knowledgeable experts who were enthusiastic about the field and forthcoming about their experiences and findings. The idea-sharing atmosphere was a perfect fit for our commitment to open source solutions.
Mike also had one final insight about the future of AI, machine learning, and the individuals who will shape the future in these fields.
Going back to Swamidass’ piano analogy, the suggestion was that a preeminent classical pianist may represent an engineer who has developed true expertise in the field of machine learning. However, Mike said he likes to think of a jazz pianist as the ultimate virtuoso.
As explained by Mike, a classical pianist may be a master at playing anything that has been written before. On the other hand, a jazz pianist knows how to utilize the tools available – technical skill and understanding of music theory – to take a theme and turn it into something new and inspired.
In Mike’s opinion, truly gifted machine learning engineers are more like jazz pianists. It’s not enough to have the skills to simply build what’s been built before. Ultimately, development of leading-edge technology solutions requires us to improvise, enabling the creation of brand-new models that change the way we experience the world and opening our minds to new possibilities.
Mike Martinez is a Principal Software Engineer for OCI. He has extensive experience defining architectures and implementing complex applications for clients in the financial services industry. He offers expertise in numerous cloud platforms, and he has presented talks on data visualization, machine learning, and R language.
To learn more about how our team may be able to help your organization take greater advantage of machine learning and AI, please contact us.
Keep Up With Current Events
Check out our events page to find out where our engineers are headed next!
