Bitstream-Based Neural Network for Scalable, Efficient, and Accurate Deep Learning Hardware.

While convolutional neural networks (CNNs) continue to renew state-of-the-art performance across many fields of machine learning, their hardware implementations tend to be very costly and inflexible. Neuromorphic hardware, on the other hand, targets higher efficiency but their inference accuracy lags far behind that of CNNs. To bridge the gap between deep learning and neuromorphic computing, we present bitstream-based neural network, which is both efficient and accurate as well as being flexible in terms of arithmetic precision and hardware size. Our bitstream-based neural network (called SC-CNN) is built on top of CNN but inspired by stochastic computing (SC), which uses bitstreams to represent numbers. Being based on CNN, our SC-CNN can be trained with backpropagation, ensuring very high inference accuracy. At the same time our SC-CNN is deterministic, hence repeatable, and is highly accurate and scalable even to large networks. Our experimental results demonstrate that our SC-CNN is highly accurate up to ImageNet-targeting CNNs, and improves efficiency over conventional digital designs ranging through 50-100% in operations-per-area depending on the CNN and the application scenario, while losing <1% in recognition accuracy. In addition, our SC-CNN implementations can be much more fault-tolerant than conventional digital implementations.

Cost-effective stochastic MAC circuits for deep neural networks.

Stochastic computing (SC) is a promising computing paradigm that can help address both the uncertainties of future process technology and the challenges of efficient hardware realization for deep neural networks (DNNs). However the impreciseness and long latency of SC have rendered previous SC-based DNN architectures less competitive against optimized fixed-point digital implementations, unless inference accuracy is significantly sacrificed. In this paper we propose a new SC-MAC (multiply-and-accumulate) algorithm, which is a key building block for SC-based DNNs, that is orders of magnitude more efficient and accurate than previous SC-MACs. We also show how our new SC-MAC can be extended to a vector version and used to accelerate both convolution and fully-connected layers of convolutional neural networks (CNNs) using the same hardware. Our experimental results using CNNs designed for MNIST and CIFAR-10 datasets demonstrate that not only is our SC-based CNNs more accurate and 40∼490× more energy-efficient for convolution layers than conventional SC-based ones, but ours can also achieve lower area-delay product and lower energy compared with precision-optimized fixed-point implementations without sacrificing accuracy. We also demonstrate the feasibility of our SC-based CNNs through FPGA prototypes.