CMU RI Thesis Proposal: Generalized Backpropagation

Title: Generalized Backpropagation

David Bradley

Robotics Institute

Carnegie Mellon University

Place: NSH 1507

Time: 2:00 PM

Date: 2 Oct 2008

Abstract:

Robotics problems are often ill-suited for the standard supervised learning paradigms. Consequently, complex tasks like autonomous driving through difficult off-road terrain are decomposed into a series of small subproblems which are each solved with independent modules through learning or hand-engineering. Unfortunately, this approach exacerbates the problem of data labeling, as training data sets must now be labeled for intermediate outputs where the "correct" label that will lead to optimal system performance may be hard to define. Additionally, even if each of the subproblems is solved reasonably well, the performance of the system as a whole may suffer because of the accumulation of small errors in each module. Figuring out which individual modules in a complex system are responsible for system-wide errors, and how to adjust them to improve performance, has consumed many man-hours. 

The gradient backpropagation algorithm has been integral for neural network training because it provides a way to translate errors in the final system performance into updates for each of the parameters in a complex network. Although it is commonly associated with neural networks, backpropagation can be used to train any system of differentiable modules. Backpropagation depends exclusively on labeled data, however, and fails on deep networks where the "blame" for errors becomes too diffuse. This thesis proposes a set of tools, collectively termed "generalized backpropagation", which are derived from recent research into learning in the online, semi-supervised, transductive, and multi-task settings, and address some key limitations of the backpropagation algorithm. These tools make it easier to learn networks of modules if labeled training data is limited and can be understood as ways to create informative priors for the parameters of each module from unlabeled and weakly labeled data. In the proposed work they will be demonstrated on a challenging mobile robot navigation problem.