Neuromorphic Compression of Tactile Data with Spiking Neural Networks
Year: 2025
Type: conference-paper
Abstract: Large-scale artificial sensing is an emerging field of research that seeks to augment robotic systems with humanlike sensing capabilities in terms of performance, robustness, and power efficiency. To accomplish this, the neuromorphic approach takes inspiration from neuronal spike processing. In this work, we use spiking neural networks (SNNs) to create low-dimensional latent representations of artificial tactile data and validate their information content through classification tasks. We tested this approach with two spiking tactile datasets which contain written digits (ST-MNIST, publicly available) and traced shapes (ST-Shape, introduced here for the first time). We attained high classification accuracy for both datasets (over 80 %) while achieving significant compression of the data (over 99 %). Our results indicate that SNN-produced latent variables are excellent at capturing high-level spatial features in the data (e.g. number of sides, or shape size), but still struggle with capturing temporal features (e.g. trace speed). Additionally, the spatial features were classified with nearmaximum accuracy in just the first 25 % of the trial. These results show that by mimicking biological processing, SNNs implemented on neuromorphic chips may provide a pathway to high-performance artificial sensing systems that deliver robust real-time sensory processing with low power consumption.
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7167828101