Dynamically-biased fixed-point LSTM for time series processing in AIoT edge device

Abstract

In this paper, a Dynamically-Biased Long Short-Term Memory (DB-LSTM) neural network architecture is proposed for artificial intelligence internet of things (AIoT) applications. Different from the conventional LSTM which uses static bias, DB-LSTM adjusts the cell bias dynamically based on the previous status. Hence, a DB-LSTM cell contains information of both the previous output and the current cell state. With more information, the DB-LSTM is able to achieve faster training convergence and better accuracy. Furthermore, weight quantization is performed to reduce the weights to either 1-bit or 2-bit, so that the algorithm can be implemented in portable edge device. With the same 100 epochs training setup, more than 70% loss reduction are achieved for floating 32-bit, 1-bit and 2-bit weights, respectively. The loss degradation due to weight quantization is also negligible. The performance of the proposed model is also validated with the classical air passenger forecasting problem. 0.075 loss and 94.96% accuracy are achieved with 2-bit weight when compared to the ground truth, which is comparable to full-length 32-bit weight.