From Encoding to Decoding: Advancing Spiking Neural Network Technology for Improved Learning

Abstract

Spiking Neural Networks (SNNs) process information through discrete spike events, yet most existing encoders focus solely on converting data into spikes without verifying whether the original signal can be faithfully recovered. Reconstruction matters for SNNs because spike encoding is inherently lossy: continuous-valued inputs must be discretised into binary events, and without a decoding pathway, there is no way to measure how much information is preserved by this conversion. If critical signal features are lost during encoding, the network cannot recover them regardless of its architecture or training. A reconstruction capability, therefore, separates encoding quality from network quality, revealing whether classification failures stem from a poor encoder or a poor learner. It also enables applications beyond classification, such as neuromorphic signal processing and sensory data transmission, where the encoded signal must be recovered rather than merely classified.

This thesis presents the Dynamic Neural Encoder and Decoder (DyNED), a unified encoding system that converts both static and dynamic data into spike trains while providing a matched decoding pathway for reconstruction. DyNED applies Fast Fourier Transform preprocessing followed by sigma-delta quantisation with iterative error feedback, enabling frequency-domain spike encoding of images and time-domain encoding of speech signals. A companion compressor, DyNEDc, combines run-length encoding with Huffman coding to losslessly compress the resulting binary spike trains.

DyNED is evaluated on four datasets - two conventional (CIFAR-10 natural images, Google Speech Commands) and two neuromorphic extensions (CIFAR10-DVS event streams, Spiking Speech Commands cochlear spikes) - using reconstruction metrics (SSI, SNR, $R^2$), compression ratio, and classification accuracy within convolutional and recurrent spiking architectures. Ablation studies across quantisation levels (2-512) show that spike density scales from 0.55 to 0.99 for images and 0.23 to 0.94 for speech, with reconstruction fidelity reaching 98-99%. Comparative experiments against uniform quantisation and mu-law companding baselines demonstrate that DyNED's error feedback mechanism produces up to 74% denser spike trains at low quantisation levels while maintaining superior reconstruction quality. DyNEDc achieves compression ratios of 0.904 (9.6% space saved) for images at 256 quantisation levels and 0.007 (99.3% space saved) for speech, matching or exceeding standard compression methods. Classification accuracy reaches 88.07% on CIFAR-10 and 92.79% on Speech Commands, exceeding a matched non-spiking CNN baseline operating on identical raw spectral features (76.59%) by approximately 16 percentage points. Furthermore, the framework achieves 76.3% and 72.96% on the neuromorphic CIFAR10-DVS and SSC extensions, respectively. This confirms that the same encoder generalises across both signal modalities and pre-spiked event streams.

These results establish that spike encoding and faithful reconstruction can coexist within a single framework, providing both the sparse representations that SNNs require and the quantifiable fidelity guarantees that principled encoder design demands. By making the encoder's information capacity explicit and measurable, DyNED provides researchers with a principled way to attribute classification failures to encoding rather than to learning, rather than relying on estimation. This capability accelerates the design of SNN architectures across modalities. The framework opens direct applications beyond classification: bandwidth-constrained sensor-to-network transmission via DyNEDc compression, energy-efficient edge inference through sparse spike representations, and faithful signal recovery for neuromorphic processing pipelines where the encoded signal must later be decoded.