HLSP-LSM: Enhancing image recognition performance of liquid state machines via brain-inspired hybrid long short-term...

Liquid state machines (LSMs) provide a biologically inspired computational paradigm that projects input data into high-dimensional dynamic representations. Spike-timing-dependent plasticity (STDP) is a widely used method for optimizing the weights of LSMs. However, relying on a single STDP rule often leads to synaptic saturation, resulting in unstable dynamics and suboptimal accuracy. To address these limitations, this paper proposes the hybrid long short-term plasticity liquid state machine (HLSP-LSM), a brain-inspired training framework that integrates excitatory and inhibitory STDP (E/I-STDP) with short-term depression (STD) to enhance network stability and computational performance. In this framework, E-STDP modulates excitatory synaptic weights to support pathway selection for relevant inputs and facilitate pattern recognition; I-STDP adaptively adjusts inhibitory connections to rapidly track changes in the excitatory input current; and STD dynamically modulates transient synaptic responses to prevent excessive excitatory or inhibitory activity. The HLSP-LSM achieves competitive classification accuracies of 98.05%, 98.49%, and 86.64% on the MNIST, N-MNIST, and Fashion-MNIST datasets, respectively. Moreover, on N-MNIST, it surpasses several recent neural architecture search (NAS)-based methods without complex search strategies. Hybrid plasticity enables the HLSP-LSM to establish a stable excitation-inhibition (E-I) balance in both its structural organization and dynamic activity. Our framework offers a scalable, interpretable, and biologically plausible approach to advancing neuromorphic computation. Our code is available at: https://github.com/LukaCityU/HLSP-LSM.