Introduction:

The last few years have seen multiple major successes of deep learning. From beating professionals at games like Go, to fast detection of cancer, classification of complex images, and generation of captions for images with incomplete information. In classification and reinforcement learning, deep learning indeed has outperformed all existing machine learning and model-based methods. Fueled by rapid advances in parallel computing and GPUs, new impressive results are achieved almost on a daily basis by deep learning methods in a vast array of applications from finance to self-driving cars. In the midst of this excitement however, there are fundamental questions that remain unanswered: For example, recent results have also shown that deep learning can be highly susceptible to noise and can misclassify images with very high confidence that are very easy for humans to process. If deep learning is going to make its way into self-driving cars and other autonomous systems, it is justifiable to ask just how stable and robust can deep learning based control and decision making algorithms be? Indeed, one could ask a even deeper question: What is the role of feedback in an increasingly data-driven world? This we believe is a question that the controls community can significantly contribute to answering. In the view of this vision, the objective of this workshop is to:

  • Introduce the theory, algorithms, and implementation of Deep Learning to the audience, including conducting tutorials on using deep learning software (Google Tensor Flow). Embedded in this introduction, is a compact introduction to machine learning in general
  • Have a candid conversation about the connection of deep learning with the Neural Network theory that has been developed in the controls community in the late 90s-early 2000s, and the larger controls efforts in system identification, adaptive control, and Kalman filtering
  • Bring together experts from deep learning and machine learning with control theory experts
  • Hold a discussions on what critical roles feedback control and estimation play in a world dominated by data-driven learning

Workshop outcomes:

The workshop is being organized by experts who have a track record of publications in control theory as well as highly competitive machine learning venues. Leveraging our contacts (and the location of the conference), the organizers will invite and bring together experts from Machine Learning to CDC if the workshop proposal is accepted. Many of these folks would typically not visit CDC, we hope that this won’t be their last CDC.

On the other hand, the workshop will also introduce deep learning to the CDC audience, and reveal to a large extent that while the fundamental concepts remain rooted in backpropagation as was studied in the controls community, the recent implementations differentiate and innovate in unique and powerful ways. To ensure that we leave our audience with a strong place to begin their journey in deep learning, we will execute two hour-long tutorials complete with sample code: 1. Classification with Google Tensor Flow package, 2. Deep Reinforcement Learning with Keras RL.

Audience

The target audience for this workshop are control theorists and practicing controls engineers. We will assume only familiarity with MATLAB and some familiarity with Python programming language (we will circulate material beforehand to the registrants to prime them in Python). Our presentations will be self-contained and leverage material that instructors have developed in teaching courses on related topics UIUC or their organizations. In addition, we will provide a repository of relevant links that help the audience navigate the wide variety of literature and resources on the internet on deep learning.

Time Title Presenter
8:00 - 8:20 AM Introduction and Welcome. Girish Chowdhary
8:20 - 8:40 AM Some results with Deep Learning Alex Scwhing
8:40 - 9:00 AM What's the big deal, economically? Chinmay Soman
9:00 - 11:00 AM Module 1: Deep learning algorithms and software. Girish Chowdhary, Alex Schwing, Hassan Kingravi, Moitreya Chatterji
11:00 - 12:00 PM Invited Talk:
Deep Neural Networks:
Function Approximation and Classification.
R.Srikant (UIUC)
12:00 - 1:00 PM Lunch
1:00 - 2:30 PM Module 2: Part 1:
Deep Reinforcement Learning.
Girish Chowdhary, Alex Schwing, Anay Pattanaik, Moitreya Chatterji, Girish Joshi
2:30 - 3:00 PM Coffe break
Coffee break
coffee break
3:00 - 4:00 PM Module 3:
Deep Recurrent Networks and Evolving Gaussian Processes
Hassan Kingravi, Girish Chowdhary, Moitreya Chatterji, Joshua Whitman
4:00 - 5:00 PM Discussion about open problems and challenges facing the community. Everyone