Stochastic Computing Architectures for Lightweight LSTM Neural Networks

For emerging edge and near-sensor systems to perform hard classification tasks locally, they must avoid costly communication with the cloud. This requires the use of compact classifiers such as recurrent neural networks of the long short term memory (LSTM) type, as well as a low-area hardware technology such as stochastic computing (SC). We study the benefits and costs of applying SC to LSTM design. We consider a design space spanned by fully binary (non-stochastic), fully stochastic, and several hybrid (mixed) LSTM architectures, and design and simulate examples of each. Using standard classification benchmarks, we show that area and power can be reduced up to 47% and 86% respectively with little or no impact on classification accuracy. We demonstrate that fully stochastic LSTMs can deliver acceptable accuracy despite accumulated errors. Our results also suggest that ReLU is preferable to tanh as an activation function in stochastic LSTMs