Exploring the Optimal Learning Technique for IBM Neurosynaptic System to Overcome Quantization Loss

Inspired by the fact that human brain is much more efficient than any nowadays computers, neuromorphic computing is aim at performing near human brain ability of processing huge amount of data in an extreme short time. For the hardware part, neuromorphic computing is also extended to systems facilitating the computation of neural network and machine learning algorithms. Recently, IBM Neurosynaptic system is one of the well-known project that dedicated on energy-efficient neural network applications. However, However, one of the known issues in TrueNorth design is the limited precision of synaptic weights, each of which can be selected from only four integers. To improve the computation accuracy and reduce the incurred hardware cost, in this work, we investigate seven different regularization functions in the cost function of the learning process on TrueNorth platform. Our experimental results proved that the proposed techniques considerably improve the computation accuracy of TrueNorth platform and reduce the incurred hardware and performance overheads.