Hyperdimensional Decoding of Spiking Neural Networks

While SNNs have the potential to replace ANNs, partly due to the prospect for significant energy savings [2], SNNs are less commonly employed as they tend to achieve lower accuracy than ANNs in several learning tasks [3], the energy efficiency of SNNs is only realized on specialised hardware [4], and implementing SNNs is more complex than ANNs as SNNs require knowledge of both neuroscience and machine learning [5]. Neuromorphic algorithms are a key facet of the maturation and wider spread adoption of SNNs. Even though SNNs are meant to be more biologically plausible than ANNs, several of the methods used by SNNs deviate from known biological mechanisms. This includes training using backpropagation through time (BPTT), or simply ignoring the temporal aspect of activity. This paper focuses on improving concept representations and SNN output decoding by developing SNN specific methodology.