[Paper Review] Synaptic Plasticity Dynamics for Deep Continuous Local Learning.

Jacques Kaiser, Hesham Mostafa+1 more|arXiv (Cornell University)|Nov 27, 2018

Advanced Memory and Neural ComputingEngineering25 citations

TL;DR

This paper proposes Deep Continuous Local Learning (DECOLLE), a biologically inspired learning rule for spiking neural networks that enables deep, temporal representation learning using only local synaptic plasticity and synthetic gradients. By deriving plasticity rules from user-defined cost functions via autodifferentiation, DECOLLE achieves performance on par with state-of-the-art methods on MNIST and DvsGesture, enabling low-power, event-based vision with high temporal precision.

ABSTRACT

A growing body of work underlines striking similarities between biological neural networks and recurrent, binary neural networks. A relatively smaller body of work, however, discusses similarities between learning dynamics employed in deep artificial neural networks and synaptic plasticity in spiking neural networks. The challenge preventing this is largely caused by the discrepancy between the dynamical properties of synaptic plasticity and the requirements for gradient backpropagation. Learning algorithms that approximate gradient backpropagation using locally synthesized gradients can overcome this challenge. Here, we show that synthetic gradients enable the derivation of Deep Continuous Local Learning (DECOLLE) in spiking neural networks. DECOLLE is capable of learning deep spatio-temporal representations from spikes relying solely on local information. Synaptic plasticity rules are derived systematically from user-defined cost functions and neural dynamics by leveraging existing autodifferentiation methods of machine learning frameworks. We benchmark our approach on the MNIST and the event-based neuromorphic DvsGesture dataset, on which DECOLLE performs comparably to the state-of-the-art. DECOLLE networks provide continuously learning machines that are relevant to biology and supportive of event-based, low-power computer vision architectures matching the accuracies of conventional computers on tasks where temporal precision and speed are essential.

Motivation & Objective

• To bridge the gap between artificial neural network training and biological synaptic plasticity by enabling local, continuous learning in deep spiking networks.
• To overcome the incompatibility between gradient backpropagation and local synaptic dynamics in spiking neural networks.
• To develop a biologically plausible learning rule that supports deep, continuous, and event-based learning in neuromorphic hardware.
• To enable high-accuracy learning on temporal and spatiotemporal tasks using only local information and synthetic gradients.

Proposed method

• Synthetic gradients are used to decouple weight updates from error backpropagation, allowing local learning in deep spiking networks.
• Plasticity rules are systematically derived from user-defined cost functions using autodifferentiation from machine learning frameworks.
• The method leverages existing automatic differentiation tools to generate biologically plausible update rules for synaptic weights.
• DECOLLE operates in a continuous, online learning regime, updating weights in real time based on local error signals.
• The approach is applied to deep spiking neural networks trained on spike-based inputs, enabling temporal representation learning.
• The framework is evaluated on standard benchmarks, including MNIST and the event-based DvsGesture dataset.

Experimental results

Research questions

• RQ1 Can synthetic gradients enable local, continuous learning in deep spiking neural networks without full backpropagation?
• RQ2 Can plasticity rules derived from cost functions via autodifferentiation produce biologically plausible and effective learning dynamics?
• RQ3 Does DECOLLE achieve competitive performance on spatiotemporal tasks like gesture recognition using event-based data?
• RQ4 Can DECOLLE support low-power, event-driven inference while maintaining accuracy comparable to conventional deep learning models?

Key findings

• DECOLLE achieves performance comparable to state-of-the-art methods on the MNIST dataset using only local learning and synthetic gradients.
• On the event-based DvsGesture dataset, DECOLLE matches the accuracy of leading models while relying solely on local synaptic plasticity.
• The method enables continuous, online learning in deep spiking networks, supporting real-time adaptation to temporal input sequences.
• Synaptic plasticity rules derived via autodifferentiation effectively approximate gradient-based learning without requiring global error signals.
• DECOLLE demonstrates feasibility for low-power, neuromorphic computing architectures that prioritize temporal precision and energy efficiency.