Hyperdimensional Decoding of Spiking Neural Networks

TL;DR

This work addresses the challenge of decoding Spiking Neural Networks (SNNs) with low latency and energy while maintaining high accuracy and robustness. It introduces SNN-HDC, which directly outputs hypervectors from SNNs and uses Hyperdimensional Computing (HDC) to classify via hypervector comparisons, enabling continuous, event-driven inference. The approach achieves energy reductions of 1.24x–3.67x on DvsGesture and 1.38x–2.27x on SL-Animals-DVS, maintains competitive accuracy, and uniquely enables unknown-class detection (e.g., 100% for unseen classes in DvsGesture). This demonstrates a practical, hardware-friendly decoding paradigm with potential for scalable, online neuromorphic applications and avenues for future enhancements like dynamic hypervectors and memory-optimized architectures.

Abstract

This work presents a novel spiking neural network (SNN) decoding method, combining SNNs with Hyperdimensional computing (HDC). The goal is to create a decoding method with high accuracy, high noise robustness, low latency and low energy usage. Compared to analogous architectures decoded with existing approaches, the presented SNN-HDC model attains generally better classification accuracy, lower classification latency and lower estimated energy consumption on multiple test cases from literature. The SNN-HDC achieved estimated energy consumption reductions ranging from 1.24x to 3.67x on the DvsGesture dataset and from 1.38x to 2.27x on the SL-Animals-DVS dataset. The presented decoding method can also efficiently identify unknown classes it has not been trained on. In the DvsGesture dataset the SNN-HDC model can identify 100% of samples from an unseen/untrained class. Given the numerous benefits shown and discussed in this paper, this decoding method represents a very compelling alternative to both rate and latency decoding.

Figures (8)

• Figure 1: Two random binary hypervectors with $10$ dimensions.
• Figure 2: Number of classes that Binary Hypervectors and One-Hot Encoding can represent given a number of dimensions.
• Figure 3: A rate decoded SNN (A) and the SNN-HDC model (B). The rate decoded SNN has one output neuron per classification. Inputs are classified based on highest spike count. Here the input is classified as Class 1. The SNN-HDC has an arbitrary amount of output neurons used to build a hypervector over time. Every dimension initialises with a value of $0$. This value is changed to $1$ if that dimension observes the presence of spikes.
• Figure 4: SNN-HDC results trained on the DvsGesture dataset amir_low_2017 for various dimensionality and membrane potential decay rate ($\beta$) values.
• Figure 5: Rate and latency decoded results trained on the DvsGesture dataset amir_low_2017 for various membrane potential decay rates ($\beta$).