Debugging ML via Feature-guided Analysis: from Neuron-based to Ensembles-based Rules

Neural networks support software engineers in their tasks and activities. For example, in the automotive domain, neural networks support automotive software by detecting other vehicles and their distance from the ego vehicle. Unlike other software systems that the engineers manually program, the neural network’s behavior is defined by its training data. Since the behavior of a neural network is learned from data, the interpretation of the reasoning procedure employed by the neural networks is complicated. This project aims to help software engineers debug their neural networks. It is framed in the context of feature-guided analysis: A recent approach to explaining neural network reasoning. The project aims to to (a) understand whether and how feature-guided analysis can help debug neural networks, and (b) extend it to provide consider neurons’ ensembles that can give engineers higher-level debugging information. The results of this work will enhance current software development practices by proposing a novel debugging technique industries can use to understand the behavior of their neural networks.

Faculty Supervisor:

Mark Lawford

Student:

Partner:

University of Bergamo

Discipline:

Engineering

Sector:

Education

University:

McMaster University

Program: