SuperSpike: Supervised learning in multi-layer spiking neural networks (1705.11146v2)

Abstract: A vast majority of computation in the brain is performed by spiking neural networks. Despite the ubiquity of such spiking, we currently lack an understanding of how biological spiking neural circuits learn and compute in-vivo, as well as how we can instantiate such capabilities in artificial spiking circuits in-silico. Here we revisit the problem of supervised learning in temporally coding multi-layer spiking neural networks. First, by using a surrogate gradient approach, we derive SuperSpike, a nonlinear voltage-based three factor learning rule capable of training multi-layer networks of deterministic integrate-and-fire neurons to perform nonlinear computations on spatiotemporal spike patterns. Second, inspired by recent results on feedback alignment, we compare the performance of our learning rule under different credit assignment strategies for propagating output errors to hidden units. Specifically, we test uniform, symmetric and random feedback, finding that simpler tasks can be solved with any type of feedback, while more complex tasks require symmetric feedback. In summary, our results open the door to obtaining a better scientific understanding of learning and computation in spiking neural networks by advancing our ability to train them to solve nonlinear problems involving transformations between different spatiotemporal spike-time patterns.

• The paper introduces SuperSpike, a surrogate gradient method that effectively trains multi-layer spiking neural networks for complex spatiotemporal tasks.
• It demonstrates that symmetric feedback in error propagation yields superior performance compared to random or uniform strategies in challenging nonlinear tasks.
• The study integrates biologically plausible mechanisms, such as nonlinear voltage dynamics and synaptic eligibility traces, to bridge computational neuroscience and AI.

SuperSpike: Supervised Learning in Multi-Layer Spiking Neural Networks

The research paper, "SuperSpike: Supervised learning in multi-layer spiking neural networks," addresses the challenge of training spiking neural networks (SNNs) for performing complex nonlinear computations on spatiotemporal spike patterns. The authors, Friedemann Zenke and Surya Ganguli, present a novel supervised learning rule, termed SuperSpike, which leverages a surrogate gradient approach to enable effective training of deterministic integrate-and-fire neurons organized in multilayer networks.

Key Contributions

1. Surrogate Gradient Approach: The paper introduces SuperSpike, a three-factor learning rule tailored for deterministic leaky integrate-and-fire (LIF) neurons. This rule is grounded in voltage-based nonlinear computations, allowing the training of SNNs on complex tasks without relying on stochastic gradient approximations.
2. Credit Assignment Strategies: The paper explores various feedback mechanisms for error propagation in hidden units, including symmetric, random, and uniform feedback, inspired by the concept of feedback alignment. It establishes that for simpler tasks, these strategies perform similarly, but symmetric feedback proves more effective for complex functions.
3. Biological Plausibility: The learning rule integrates several biologically plausible components like nonlinear voltage-dependent terms and synaptic eligibility traces, drawing parallels with spike-timing-dependent plasticity (STDP) observed in biological neurons.

Experimental Results

The authors validate the efficacy of the SuperSpike rule through numerical experiments. On relatively simple tasks, such as precise spike timing and XOR problem-solving, networks using SuperSpike exhibit successful learning across various feedback types; however, symmetric feedback results in more reliable performance on challenging tasks involving extensive spatiotemporal transformations.

Key results include:

• Performance comparable or superior to traditional gradient approaches, with symmetric feedback achieving high accuracy on tasks that require transformation between complex spike-time patterns.
• In scenarios with random feedback, the introduction of heterosynaptic regularization aids in controlling firing rates and improves convergence, though not to the level of symmetric feedback.

Practical and Theoretical Implications

Practically, SuperSpike enhances our ability to train SNNs, which are crucial for developing neuromorphic computing hardware that can emulate human brain-like calculations efficiently. Theoretically, it offers insights into potential mechanisms of learning and credit assignment in biological neural networks, contributing to the broader understanding of in-vivo learning dynamics.

Future Directions

Further exploration into intelligent neuromodulation or biologically inspired mechanisms remains open. Extending this framework to deep and recurrent SNNs offers a promising research avenue, potentially bolstering their application in real-world AI systems and shedding light on the neural basis of temporal computations.

By developing SuperSpike, this paper advances the field's understanding of supervised learning in temporally coding SNNs, contributing significantly to both computational neuroscience and artificial intelligence domains.