TL;DR
This paper introduces a novel online learning algorithm for Spiking Neural Networks (SNNs) that leverages multiple "compartments" within each neuron to improve the accuracy of learning. By using these multiple samples, the algorithm achieves better log-likelihood estimation and gradient accuracy, leading to improved performance on neuromorphic tasks.
What Problem Does This Paper Address?
Spiking Neural Networks (SNNs) are a promising alternative to traditional Artificial Neural Networks (ANNs) due to their potential for significantly lower energy consumption, mimicking the efficiency of biological brains. However, training SNNs is challenging because their discrete, binary spiking outputs make gradient-based learning difficult. Existing methods often rely on single samples to estimate gradients, which can lead to inaccurate learning signals. This paper addresses the challenge of improving the accuracy and efficiency of SNN training by proposing a method that utilizes multiple independent spiking signals from within each neuron to obtain more robust statistical estimates for learning.
Methods
The core innovation is a Multi-Compartment SNN model. In this model, each spiking neuron is equipped with K independent "compartments." Each compartment tracks its own membrane potential and generates its own spiking signal, but all compartments within a neuron share the same synaptic weights. This architecture is combined with a Generalized Expectation-Maximization (GEM) approach, specifically a variant called GEM-VLSNN (Generalized EM Variational online Learning for SNNs).
GEM-VLSNN uses importance sampling with the K independent spiking samples from the compartments to approximate the log-likelihood bound more accurately. This multi-sample approach provides a better statistical estimate of the log-likelihood and its gradient compared to single-sample methods. The learning rule is an online, local update rule, meaning that each neuron can update its weights based on local information and a global learning signal, making it suitable for neuromorphic hardware.
Datasets
The paper evaluates its proposed method on the neuromorphic MNIST-DVS dataset. This dataset consists of event-based camera recordings of handwritten digits. The authors focus on a classification task for three digits ({0, 1, 2}).
Key Insights and Ideas
- Compartments as Statistical Power: The key insight is that multiple independent spiking outputs from within a single neuron (via compartments) can be treated as multiple samples. These samples, when used together, provide a richer statistical picture for learning than a single sample.
- GEM for Better Bounds: The Generalized Expectation-Maximization (GEM) framework is crucial for optimizing a tighter approximation of the log-likelihood, which is essential for effective learning in SNNs.
- Online and Local Learning: The GEM-VLSNN algorithm is designed for online learning, allowing for continuous updates as data arrives. It also employs local learning rules, which are highly desirable for efficient implementation on neuromorphic hardware.
- Trade-off between Performance and Communication: The number of compartments (K) offers a flexible trade-off. Increasing K improves testing performance (log-likelihood, accuracy, calibration) but also increases the communication load required to broadcast the global learning signal.
Key Results
- Improved Log-Likelihood and Accuracy: Experiments show that increasing the number of compartments (K) during training leads to significant improvements in both estimated log-likelihood and classification accuracy on the MNIST-DVS dataset. For instance, increasing K from 1 to 20 yielded better results.
- Enhanced Calibration: The proposed method also improves the calibration of the SNNs, meaning the network's confidence in its predictions better reflects its actual accuracy.
- Linear Increase in Communication Load: The broadcast communication load from the central processor to neurons increases linearly with K, highlighting the trade-off between performance and communication overhead.
Limitations and Future Work
- Communication Overhead: The paper acknowledges that increasing the number of compartments (K) leads to a linear increase in communication load, which could be a bottleneck in highly distributed systems.
- Future Work: The authors suggest extending the learning algorithm to other reward functions beyond log-likelihood, such as the Van Rossum distance, and exploring its application to networks with more complex spiking mechanisms like Winner-Take-All (WTA) circuits that process multi-valued spikes.
Summary updated