Stock price prediction based on an energy-efficient spiking-LSTM hardware accelerator

Abstract

Inspired by the way the human brain thinks, the neuromorphic system applies the principles of biological brains to computer architecture, providing low-energy, distributed, and massively parallel advantages for brain-inspired systems. This work presents an energy-efficient spiking long short-term memory (sLSTM) neural network hardware accelerator for sequence prediction applications, containing 256 neurons and 64k synapses in 0.96 mm2 area. The sLSTM model can process time-dependent data and realize long-term and short-term memory to forget, memorize selectively. A leaky integrate and fire (LIF) neuron model is proposed to characterize the stimulation of neuronal membrane potentials using simple digital logic circuit without any multipliers, which extremely reduces the power consumption of the hardware system. Accordingly, the chip achieved an energy efficiency of 10.3 uJ@50 MHz per sample and a predicting accuracy of about 93.2% in sLSTM neural network model using the stock price of Google from Yahoo finance, based on the modified LIF neuron.