Comparing and Contrasting Artificial Neural Networks with Biological Neural Networks for Improved Representation Learning

Computer Sciences; Computational Neuroscience; Artificial Intelligence/Robotics; Philosophy

We will study what are the implications of the different modes of operation of artificial neural networks and biological neural networks, and how do they build abstraction into their representations.

Research Interests
  • Computer Vision
  • Neuroscience
  • Philosophy & Humanities
  1. Vicente Ordonez-Roman headshot
    Vicente Ordonez-Roman
    School of Engineering and Applied Science
  2. Paul Humphreys headshot
    Paul W. Humphreys
    College of Arts and Sciences
  3. William Levy headshot
    William B Levy
    School of Medicine

Deep neural networks have become pervasive models in applications ranging from smart voice assistive agents, to autonomous navigation agents. Understanding the representations learned by deep neural networks is thus of great practical importance. The type of abstraction or compression performed by these models beyond the first layer is currently poorly understood and not well defined. Despite this, it is widely held that many of these representations are compositional and involve successive processes of abstraction. Our goal, by leveraging concepts from artificial intelligence, neuroscience, and philosophy, is to uncover from a quantitative perspective some measures of abstraction, the relation of abstraction to generalization, and what the barriers are to human understanding of the internal representations. Neural computation in the mammalian brain is dominated by feedback more than feedforward computation. However, computation in conventional deep AI-networks is all feedforward. We conjecture that feedback mechanisms provide the technical ability to freely mix and match sets of microscopic abstractions encoded by small groups of neurons where such groups are characteristic of cortical columns and as assembled into defined cortical regions. Understanding these mechanisms will also allow for more sophisticated designs of artificial neural networks.

Desired outcomes

We would like to obtain the following outcomes:
* More nuanced understanding of the representations encoded in the intermediate activations of neural networks.
* More nuanced understanding of the dynamic processes executed in neural networks (feed-forward vs feed-back based)
* Actionable information to build more robust architectures, representations, and inference procedures.