MMT-SNN: Markovian decision and multi-threshold spike delivery integrated adaptive spiking neural network for tactile object recognition

Abstract

Tactile object recognition represents a significant research direction within the field of robotic perception. Traditional frame-based tactile object recognition methods encounter limitations when applied to event-driven tactile data, especially under rapid dynamic change conditions. In contrast, spiking neural networks (SNNs) demonstrate higher efficiency when processing event-driven tactile data streams. However, the parameter update strategies employed by the existing SNN models typically rely on fixed learning strategies and regularization parameters, which may lead to slow convergence or entrapment in local optima when handling complex, variable tactile signals. Moreover, the current SNN models for tactile recognition often utilize single-neuron firing mechanisms, which restricts their overall neuronal expression capacities. To address these issues, we propose the MMT-SNN method, which leverages Markov decision process (MDP) principles to dynamically adjust the parameter update strategies of SNNs, thereby enhancing the accuracy and efficiency of object recognition. Additionally, a multi-threshold firing mechanism is employed to attain improved gradient propagation and increased neuronal expressiveness within the network. Experimental results demonstrate that MMT-SNN significantly outperforms the state-of-the-art approaches, achieving a 12.50% performance improvement over the classic TactileSGNet approach on the Containers-v0 dataset and a 3.61% improvement on the Objects-v0 dataset.