Ponte: Represent Totally Binary Neural Network Toward Efficiency

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Binary Neural Networks (BNNs) offer computational efficiency but traditionally use full-precision first and last layers.
• This conventional approach increases logic usage in Field-Programmable Gate Array (FPGA) implementations.

Purpose of the Study:

• To develop a novel approach, Ponte, for extending binarization to the first and last layers of BNNs.
• To mitigate computational overhead and logic usage in FPGAs without compromising network accuracy.

Main Methods:

• Ponte extends binarization to all network layers, including the first and last.
• Employs Ponte::encoding for unique data representation and Ponte::dispatch/Ponte::sharing for channel duplication strategies.
• All methods are back-propagation supported, enabling implementation and training.

Main Results:

• Ponte successfully binarizes all layers, preserving input data integrity.
• The approach enhances the representational capacity of BNNs.
• Achieved comparable or superior performance metrics on CIFAR-10 and ImageNet datasets with reduced computational demands.

Conclusions:

• Ponte represents a significant advancement in creating fully binary neural networks.
• The method facilitates practical deployment of BNNs in resource-constrained environments.
• Demonstrates the feasibility and effectiveness of fully binarized networks through extensive experimentation.