A New Focus on Deep Learning

By Jeremy Watt

For eons, humans have sought out rules, often in the form of mathematical formulae, that describe how important systems in the world around us work. We do this because if we can learn mathematical patterns that describe how a system works, whether it be biological, mechanical or financial, we not only can better understand it, but we can predict its future behavior and – ultimately – control it.  

However, the process of finding the right formula, one that closely describes a given phenomenon, has historically been no easy task. For most of our history, data has been a scarce commodity, and our ability to compute has been limited to what we could accomplish by hand. Both of these factors naturally limited the range of systems we could investigate. It forced us to use more philosophical approaches to rule-finding.

But today, we live in a world awash in data, and we have colossal computing power at our fingertips. Because of this, we can tackle a much wider array of systems and take a much more empirical approach to rule-finding.

‘Machine learning’ is the name we use to denote the broad (and growing) collection of modern pattern-finding algorithms designed to properly identify system rules empirically by leveraging enormous amounts of data and computing power. Deep learning is a powerful subset of these algorithms that work especially well with structured data like images, videos and time series. This inherent strength has made them the modern de-facto choice when modeling systems in computer vision, natural language processing, autonomous systems and robotics, among other areas.

This past school year, I had the opportunity to pilot a series of three new deep learning courses within Northwestern Engineering. Our overarching goal with all three courses — Deep Learning Foundations From Scratch, Deep Reinforcement Learning From Scratch and Optimization Techniques for Machine Learning and Deep Learning —  was to help students gain a rock-solid foundational understanding of how the tools of deep learning and reinforcement learning work, as well as when and how they should be applied. We do this by providing lectures that communicate the complex technical details of these subjects as clearly as possible, employing a variety of visualizations and examples that frame the intuition behind big ideas and formulae.  

We also do this by making our students work, with homework assignments that require extensive coding, since we believe that ‘getting your hands dirty’ is the best way to gain technical mastery in any discipline.

Looking back on the first year of the courses, I think all three went quite well. Deep learning and reinforcement learning provides us with a range of fundamental components of ‘intelligence’. If we leverage them properly, we can build smarter robots that perform complex tasks automatically.  

Machine learning has become in our time what electricity was over a century ago — a tool with endless applications. Like electric power, it will become tightly integrated into the everyday workflow of science and industry — to address a range of tasks whose breadth we cannot yet see.

Jeremy Watt is a deep learning and machine learning researcher, developer, educator, and tinkerer with a wealth of experience building AI systems and helping others build them too. He has co-authored a university-level textbook on the subject called Machine Learning Refined. He also is a co-founder of Degree Six (www.dgsix.com), a Chicago-based consulting firm providing machine learning and deep learning software development and training services to businesses large and small.