Multiplicative Coding and Factorization in Vector Symbolic Models of Cognition

Spencer Kent

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2020-215
December 18, 2020

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-215.pdf

This dissertation covers my attempts to confront the challenge and promise of multiplicative representations, and their attendant factorization problems, in the brain. This is grounded in a paradigm for modeling cognition that defines an algebra over high-dimensional vectors and presents a compelling factorization problem. The proposed solution to this problem, a recurrent neural network architecture called Resonator Networks, has several interesting properties that make it uniquely effective on this problem and may provide some principles for designing a new class of neural network models. I show some applications of multiplicative distributed codes for representing visual scenes and suggest how such representations may be a useful tool for unifying symbolic and connectionist theories of intelligence.

Advisor: Bruno Olshausen and Alexei (Alyosha) Efros


BibTeX citation:

@phdthesis{Kent:EECS-2020-215,
    Author = {Kent, Spencer},
    Title = {Multiplicative Coding and Factorization in Vector Symbolic Models of Cognition},
    School = {EECS Department, University of California, Berkeley},
    Year = {2020},
    Month = {Dec},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-215.html},
    Number = {UCB/EECS-2020-215},
    Abstract = {This dissertation covers my attempts to confront the challenge and promise of multiplicative representations, and their attendant factorization problems, in the brain. This is grounded in a paradigm for modeling cognition that defines an algebra over high-dimensional vectors and presents a compelling factorization problem. The proposed solution to this problem, a recurrent neural network architecture called Resonator Networks, has several interesting properties that make it uniquely effective on this problem and may provide some principles for designing a new class of neural network models. I show some applications of multiplicative distributed codes for representing visual scenes and suggest how such representations may be a useful tool for unifying symbolic and connectionist theories of intelligence.}
}

EndNote citation:

%0 Thesis
%A Kent, Spencer
%T Multiplicative Coding and Factorization in Vector Symbolic Models of Cognition
%I EECS Department, University of California, Berkeley
%D 2020
%8 December 18
%@ UCB/EECS-2020-215
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-215.html
%F Kent:EECS-2020-215