Abstract
Liquid State Machine (LSM) is a brain-inspired architecture used for solvingproblems like speech recognition and time series prediction. LSM comprises of arandomly connected recurrent network of spiking neurons. This networkpropagates the non-linear neuronal and synaptic dynamics. Maass et al. haveargued that the non-linear dynamics of LSMs is essential for its performance asa universal computer. Lyapunov exponent (mu), used to characterize the"non-linearity" of the network, correlates well with LSM performance. Wepropose a complementary approach of approximating the LSM dynamics with alinear state space representation. The spike rates from this model are wellcorrelated to the spike rates from LSM. Such equivalence allows the extractionof a "memory" metric (tau_M) from the state transition matrix. tau_M displayshigh correlation with performance. Further, high tau_M system require lesserepochs to achieve a given accuracy. Being computationally cheap (1800x timeefficient compared to LSM), the tau_M metric enables exploration of the vastparameter design space. We observe that the performance correlation of thetau_M surpasses the Lyapunov exponent (mu), (2-4x improvement) in thehigh-performance regime over multiple datasets. In fact, while mu increasesmonotonically with network activity, the performance reaches a maxima at aspecific activity described in literature as the "edge of chaos". On the otherhand, tau_M remains correlated with LSM performance even as mu increasesmonotonically. Hence, tau_M captures the useful memory of network activity thatenables LSM performance. It also enables rapid design space exploration andfine-tuning of LSM parameters for high performance.