CMU PhD Oral: Learning in Modular Systems

CMU PhD Thesis Defense:
David M. Bradley

Title: Learning in Modular Systems

Abstract: Complex robotics systems are often built as a system of modules, where each module solves a separate data processing task to produce the complex overall behavior that is required of the robot. For instance, the perception system for autonomous off-road navigation discussed in this thesis uses a terrain classification module, a ground-plane estimation module, and a path-planning module among others. Splitting a complex task into a series of sub-problems allows human designers to engineer solutions for each sub-problem independently, and devise efficient specialized algorithms to solve them. However, modular design can also create problems for applying learning algorithms. Ideally, learning should find parameters for each module that optimize the performance of the overall system. This requires obtaining ``local'' information for each module about how changing the parameters of that module will impact the output of the system.

Previous work in modular learning showed that if the modules of system were differentiable, gradient descent could be used to provide this local information in “shallow” systems containing with two or three modules between input and output. However, except for convolutional neural networks, this procedure was rarely successful in “deep” systems of more than three modules. Many robotics applications added an additional complication by employing a planning algorithm to produce their output. This makes it hard to define a “loss” function to judge how well the system is performing, or compute a gradient with respect to previous modules in the system.

Recent advances in learning deep neural networks suggest that learning in deep systems can be successful if data-dependent regularization is first used to provide relevant local information to the modules of the system, and the modules are then jointly optimized by gradient descent. Concurrently, research in imitation learning has offered effective new ways of defining loss functions for the output of planning modules.

This thesis combines these lines of research to develop new tools for learning in modular systems. As data-dependent regularization has been shown to be critical to success in deep modular systems, several significant contributions are provided in this area. A novel, differentiable formulation of sparse coding is presented and shown to be a powerful semi-supervised learning algorithm. Sparse coding has traditionally used non-convex optimization methods, and an alternative, convex formulation is developed with a deterministic optimization procedure. Theoretical contributions developed for this convex formulation also enable an efficient, online multi-task learning algorithm. Results in domain adaptation provide further regularization options. To allow joint optimization of systems that employ planning modules, this thesis leverages loss functions developed in recent imitation learning research, and develops techniques for improving all modules of the system with subgradient descent. Finally, this thesis has also made significant contributions to mobile robot perception for navigation, providing terrain classification techniques that been incorporated into fielded industrial and government systems. [Full PDF]


Thesis Committee Members:
James A. Bagnell, Chair
Martial Hebert
Fernando De la Torre
Yoshua Bengio, University of Montreal