The Critical Dynamics in Neural Network Improve the Computational Capability of Liquid State Machines

In recent years, increasing studies have shown that the networks in the brain can reach a critical state where dynamics exhibit a mixture of synchronous and asynchronous firing activity. It has been hypothesized that the homeostatic level balanced between stability and plasticity of this critical state may be the optimal state for performing diverse neural computational tasks. Motivated by this, the role of critical state in neural computation based on liquid state machines (LSM), which is one of the neural network application model of liquid computing, has been investigated in this note. Different from a randomly connect structure in liquid component of LSM in most studies, the synaptic weights among neurons in proposed liquid are refined by spike-timing-dependent plasticity (STDP); meanwhile, the degrees of neurons excitability are regulated to maintain a low average activity level by Intrinsic Plasticity (IP). The results have shown that the network yield maximal computational performance when subjected to critical dynamical states.