Improving Performance of Spike-based Deep Q-Learning using Ternary Neurons

We propose a new ternary spiking neuron model to improve the representation capacity of binary spiking neurons in deep Q-learning. Although a ternary neuron model has recently been introduced to overcome the limited representation capacity offered by binary spiking neurons, we show that its performance is worse than that of binary models in deep Q-learning tasks, contradicting previous findings from recent studies. Through mathematical and empirical analysis, we hypothesize that gradient estimation bias during training is the underlying cause. The proposed ternary spiking neuron model mitigates this issue by reducing the estimation bias. We use the proposed ternary spiking neuron as the fundamental computing unit in a deep spiking Q-learning network, which we call the deep asymmetric ternary spiking Q-network (DATSQN), and evaluate the network's performance in seven Atari games from the Gym environment. The results show that the proposed ternary spiking neuron mitigates the performance degradation of ternary neurons in DQN tasks and improves the mean game score relative to the binary baseline under the evaluation settings used in this paper.