A Sparsity-Driven Backpropagation-Less Learning Framework Using Populations of Spiking Growth Transform Neurons

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuromorphic Engineering

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation but face challenges in training and controlling network dynamics.
• Growth-Transform (GT) neurons provide a framework for independent control of spiking statistics and population dynamics.
• Existing training methods often rely on backpropagation, which is biologically implausible and computationally intensive.

Purpose of the Study:

• To develop a backpropagation-less learning approach for training spiking GT neuron networks.
• To enforce sparsity constraints on network spiking activity for energy efficiency.
• To enable resource-constrained neuromorphic systems like tinyML.

Main Methods:

• A novel learning framework trains spiking GT neurons by enforcing sparsity constraints on network activity.
• Spike responses are generated as a result of constraint violation, acting as Lagrangian parameters.
• The method utilizes local, online learning rules and allows for incorporation of structural constraints.

Main Results:

• The proposed framework enables networks to learn optimal parameters using biologically plausible local learning rules.
• Networks trained with this method minimize spiking activity (sparsity) and operate with maximum dynamic range.
• Demonstrated comparable classification accuracy to standard methods on a machine olfaction dataset, showcasing resource-efficient learning.

Conclusions:

• The developed learning approach is effective for training spiking GT neuron networks in an energy-efficient manner.
• This method is suitable for designing neuromorphic tinyML systems with limited resources.
• The framework successfully balances network sparsity, dynamic range, and task performance.