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

Area of Science:

• Machine Learning
• Computer Engineering
• Artificial Intelligence Hardware

Background:

• Convolutional Neural Networks (CNNs) achieve state-of-the-art performance but require costly and inflexible hardware.
• Neuromorphic hardware offers efficiency but suffers from lower inference accuracy compared to CNNs.

Purpose of the Study:

• To bridge the gap between deep learning and neuromorphic computing by developing an efficient, accurate, and flexible neural network architecture.
• To present a bitstream-based neural network (SC-CNN) that leverages stochastic computing principles.

Main Methods:

• Developed SC-CNN, a novel architecture built upon CNN principles but incorporating stochastic computing (SC) using bitstreams for number representation.
• Trained SC-CNN using backpropagation to ensure high inference accuracy, maintaining deterministic and repeatable results.
• Evaluated SC-CNN's performance, accuracy, and efficiency against conventional digital designs and CNNs targeting datasets like ImageNet.

Main Results:

• SC-CNN demonstrated high inference accuracy comparable to ImageNet-targeting CNNs.
• Achieved 50-100% improvement in operations-per-area efficiency over conventional digital designs, with less than 1% loss in recognition accuracy.
• SC-CNN implementations exhibited enhanced fault tolerance compared to traditional digital designs.

Conclusions:

• SC-CNN successfully integrates the accuracy of deep learning with the efficiency of neuromorphic computing.
• The proposed bitstream-based approach offers a flexible, scalable, and cost-effective solution for hardware implementations of neural networks.
• SC-CNN presents a promising direction for developing next-generation AI hardware.