Structured representations of brain circuit algorithms vs. unstructued representations of neural network algorithms

Abstract

Brains substantially outperform artificial methods on tasks ranging from low-? level perception to “cognitive” tasks such as language, planning, situation awareness, co-? robot collaboration, and more. Despite recent promising successes of “deep” neural network systems, performance on these tasks remains far short of human abilities. One possible reason is that, although attention has been paid to “operating rules” for neural networks, i.e., the mathematical operations that individual cells (units) and synapses (connection weights) carry out, and although successively trained hierarchical layers have been increasingly used, nonetheless there are substantial missing biologically relevant features in such networks. Perhaps most notably, very few instances exist of “structure” in networks, i.e., circuit layout designs of the kind that occur in actual brain circuitry. Brain circuit structure, i.e., multiple distinct neuron types embedded in different circuit layouts, may confer powerful computational capabilities to networks. Modeling work has provided evidence that some of these anatomical structured designs lead to a range of novel and powerful algorithms (e.g., Rodriguez et al., 2004; Granger 2006; Chandrashekar et al., 2013), several of which have been carefully shown to surprisingly outperform current standard methods in head to head comparisons (e.g., Chandrashekar & Granger 2012). Moreover, these novel algorithms have been shown to generate structured representations, beyond the standard “isa” structures of all standard neural networks. In current work in progress, the resulting structured representations are being formally evaluated in terms of the mathematics of grammars, providing explanatory accounts of their representational capabilities. The three primary specific aims of the proposal are to i) study the abilities of these brain circuit algorithms to produce representations that include structure relations beyond “isa” hierarchies; ii) identify what types of non-? isa relations can be constructed (e.g., follows / followed-? by; part-? of / contains; left-? right / above-? below); and iii) study relationships between these structured representations and particular applications data (e.g., visual objects; vision-? based navigation; auditory speech sounds).

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 12, 2016
Source ID
N000141512823

Entities

People

  • Richard Granger

Organizations

  • Board of Trustees of Dartmouth College
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Neural Network Machine Learning.
  • Neuroscience

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Autonomy