Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ELiSe: Efficient Learning of Sequences in Structured Recurrent Networks

Laura Kriener, Kristin Völk, Ben von Hünerbein, Federico Benitez, Walter Senn, Mihai A. Petrovici

TL;DR

This work proposes a model that builds on the presence of a certain network scaffold at the onset of learning and the existence of dendritic compartments for enhancing neuronal information storage and computation and builds on these features to acquire and replay complex non-Markovian spatio-temporal patterns using only local, always-on and phase-free synaptic plasticity.

Abstract

Behavior can be described as a temporal sequence of actions driven by neural activity. To learn complex sequential patterns in neural networks, memories of past activities need to persist on significantly longer timescales than the relaxation times of single-neuron activity. While recurrent networks can produce such long transients, training these networks is a challenge. Learning via error propagation confers models such as FORCE, RTRL or BPTT a significant functional advantage, but at the expense of biological plausibility. While reservoir computing circumvents this issue by learning only the readout weights, it does not scale well with problem complexity. We propose that two prominent structural features of cortical networks can alleviate these issues: the presence of a certain network scaffold at the onset of learning and the existence of dendritic compartments for enhancing neuronal information storage and computation. Our resulting model for Efficient Learning of Sequences (ELiSe) builds on these features to acquire and replay complex non-Markovian spatio-temporal patterns using only local, always-on and phase-free synaptic plasticity. We showcase the capabilities of ELiSe in a mock-up of birdsong learning, and demonstrate its flexibility with respect to parametrization, as well as its robustness to external disturbances.

ELiSe: Efficient Learning of Sequences in Structured Recurrent Networks

TL;DR

This work proposes a model that builds on the presence of a certain network scaffold at the onset of learning and the existence of dendritic compartments for enhancing neuronal information storage and computation and builds on these features to acquire and replay complex non-Markovian spatio-temporal patterns using only local, always-on and phase-free synaptic plasticity.

Abstract

Behavior can be described as a temporal sequence of actions driven by neural activity. To learn complex sequential patterns in neural networks, memories of past activities need to persist on significantly longer timescales than the relaxation times of single-neuron activity. While recurrent networks can produce such long transients, training these networks is a challenge. Learning via error propagation confers models such as FORCE, RTRL or BPTT a significant functional advantage, but at the expense of biological plausibility. While reservoir computing circumvents this issue by learning only the readout weights, it does not scale well with problem complexity. We propose that two prominent structural features of cortical networks can alleviate these issues: the presence of a certain network scaffold at the onset of learning and the existence of dendritic compartments for enhancing neuronal information storage and computation. Our resulting model for Efficient Learning of Sequences (ELiSe) builds on these features to acquire and replay complex non-Markovian spatio-temporal patterns using only local, always-on and phase-free synaptic plasticity. We showcase the capabilities of ELiSe in a mock-up of birdsong learning, and demonstrate its flexibility with respect to parametrization, as well as its robustness to external disturbances.
Paper Structure (25 sections, 22 equations, 7 figures, 1 table)

This paper contains 25 sections, 22 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Model description
  3. Early development provides a scaffold for learning
  4. Neuronal dynamics and dendritic learning
  5. Results
  6. Local plasticity enables complex sequence replay
  7. Learning of latent connectivity enables sustained pattern replay
  8. Learning engenders robustness to internal variability and external disruption
  9. Discussion
  10. Comparison to other approaches
  11. Neural correlates
  12. Open questions and future work
  13. Methods
  14. Neuron model
  15. Developmental setup
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Development of somatic scaffold for the learning of dendritic connections.(a) The network composed of structured pyramidal neurons is divided into an output and latent population. (b) During development, axons (orange) extend to form a sparse scaffold of somato-somatic connections. Its structure is controlled by parameters $p$ and $q$. (c) The connections of the somatic scaffold are static and have associated delays. Inhibition in the scaffold is mediated by interneurons. (d) Following development, somato-dendritic synapses (green) are learned according to a three-factor plasticity rule.
  • Figure 2: Learning process and replay ability of the network.(a) Evolution of network activity during learning. Output neurons are first nudged towards a particular target and then released in order to observe the network’s spontaneous activity (as demarcated by the red line). Snapshots during early (left), intermediate (middle) and final (right) stages of training. (b) Performance of networks with different latent population sizes measured by the loss between the activity of the output neurons and their target activity. Top: Evolution of the error during validation intervals, i.e. first replay of the pattern after being released from external nudging) during training. Bottom: Stability of replay after training, for continuous observation over many replay cycles following release from nudging. (c) Same as (b) but using the cross-correlation between output and target activities as a performance measure.
  • Figure 3: Comparison of replay performance between networks with and without learning of latent connections.(a) Accuracy and stability of sequence replay, measured by the between generated and target activity. Top: between target and produced pattern during training. Bottom: during continued replay after training. (b) Recorded activity of the output (bottom) and latent (top) neurons of the fully plastic network during free replay of the pattern after training. The red line marks the end of the external input from the teacher. (c) Same as (b) but for the ablated network.
  • Figure 4: Robustness of learning with respect to changes in the network parameters.(a) Sweep over network parameter $p$. For each $p$ runs with a range of $q$ values and multiple seeds are averaged. (b) Sweep over network parameter $q$. For each $q$ runs with a range of $p$ vlaues and multiple seeds are averaged. (c) Training of networks with sparse somato-dendritic connectivity; $r$ represents the proportion of nonzero and plastic synapses out of all possible somato-dendritic connections in the network. (d) Same as (c) but the somato-dendritic connections that match an already existing somato-somatic connection are protected from sparsification.
  • Figure 5: Network replay is robust to strong disruptions of its output activity.(a) Left: for 10 different networks during disturbed pattern replay. Right: Activity in the output population during pattern replay (red line marks end of teacher nudging). During the second replay pattern output activity is suppressed (black box) by clamping the membrane voltages of all output neurons. (b) Same as (a) but with a disturbance for the duration of a whole pattern. (c) Short disturbance but instead of suppressing neuron activity, "wrong" activity is introduced into the network by clamping to a high membrane voltage. (d) Same as (c) but with a disturbance for the duration of a whole pattern.
  • ...and 2 more figures