A Survey of Stochastic Computing Neural Networks for Machine Learning Applications

Area of Science:

• Computer Science
• Electrical Engineering
• Artificial Intelligence

Background:

• Neural networks (NNs) demand substantial hardware and energy.
• Stochastic computing (SC) offers a potential solution for efficient NN implementation.
• SC NNs aim to balance hardware efficiency with computational performance.

Purpose of the Study:

• To survey various types of SC NNs.
• To discuss recent advancements in SC NN design.
• To illustrate the versatility of SC NNs in training and inference.

Main Methods:

• Design of a basic SC neuron.
• Survey of SC implementations for multilayer perceptrons, deep belief networks, convolutional NNs, and recurrent NNs.
• Review of recent SC design progress.

Main Results:

• SC NNs offer reduced hardware requirements and power consumption.
• Performance of SC NNs has improved, becoming comparable to binary designs.
• SC NNs demonstrate versatility for both training and inference.

Conclusions:

• SC NNs present a compelling alternative to conventional NNs for hardware efficiency.
• Further research in SC designs can unlock greater performance and efficiency.
• SC NNs face challenges and advantages when compared to binary counterparts.