Towards more biologically plausible learning in neural networks
Abstract
Our overarching goal is to push toward biologically plausible and biologically inspired learning and intelligence. We are interested in biologically plausible learning for multiple reasons: 1) Biological algorithms likely capture the heuristics that make us efficient learners and solvers of problems in the world, even if they are not exactly performing stochastic gradient descent. 2) Related to 1), they might incorporate aspects of unsupervised learning into the supervised learning stream, and thus be specifically efficient with respect to use of supervised data. 3) Besides being data-efficient, they might also be morehardware-efficient and time-efficient, because biological solutions to problems often eschew computationally greedy processes like matrix inversion and large random-access memory stores. This proposal seeks support for work to progress in the direction of that broad goal, specifically by focusing on the derivation, theory, testing and improvement of biologically plausible synaptic plasticity rules for supervised learning with unsupervised elements, in recurrent and feedforward networks. We also focus on the development of network motifs and architectures seen in the brain.Specifically, we propose to develop rules for error-driven learning which, in contrast to the standard stochastic gradient rule of backpropagation, run primarily in the forward direction, so that traces of recent activation in an eligibility trace are related to error signals to drive learning in recurrent networks. We will build on our preliminary work in this direction, proposing versions of forward-running rules for feedforward networks and spiking networks. We seek to develop a better theoretical understanding of the forward-running credit assignment problem and propose alternative forms with greater biologicalplausibility. Finally, we propose to incorporate biological motifs likemodularity, gating, and nonlinear or multiplicative dendritic processingthat enrich the functionality of individual neural units, through theuse of biological constraints and bounds that nudge networks towardssolutions that involve modularity.We expect that these lines of work will produce biologically plausible learning rules with good performance relative to the backpropagation algorithm, with increases in computational efficiency and potentially sample efficiency. Given the ubiquitous use of deep neural networks in various technological, modeling, and scientific applications, and the need for greater efficiency in such algorithms if they are to be deployed more globally including on mobile platforms, as well asthe neuroscientific imperative that we cannot understand the brainwithout understanding its learning rules, we believe that this researchapproach has high potential impact for DoD capabilities, society, andscience.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2019
- Source ID
- N000141912584
Entities
People
- Ila R Fiete
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy