Synaptic Plasticity Dynamics for Deep Continuous Local Learning (DECOLLE).

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Biological neural networks and artificial neural networks share similarities.
• Bridging learning dynamics between deep artificial neural networks and spiking neural networks is challenging due to gradient backpropagation requirements.
• Existing methods struggle with the discrepancy between synaptic plasticity dynamics and gradient backpropagation.

Purpose of the Study:

• Introduce Deep Continuous Local Learning (DECOLLE), a novel spiking neural network.
• Enable online learning in spiking neural networks using local error functions.
• Develop a biologically plausible and hardware-compatible learning framework.

Main Methods:

• DECOLLE utilizes local error functions for online learning without memory overhead for gradient computation.
• Synaptic plasticity rules are systematically derived from cost functions and neural dynamics using autodifferentiation.
• The approach leverages existing machine learning frameworks for gradient approximation.

Main Results:

• DECOLLE learns deep spatio-temporal representations from spikes using only local information.
• Performance on N-MNIST and DvsGesture datasets is comparable to state-of-the-art.
• Networks demonstrate continuous learning capabilities.

Conclusions:

• DECOLLE offers a biologically relevant and neuromorphic hardware-compatible learning approach.
• It supports event-based, low-power computer vision with high temporal precision and speed.
• DECOLLE achieves accuracies matching conventional computers on specific tasks.