New Framework Boosts Binarized Neural Network Pruning Efficiency

Deep neural networks can be dramatically compressed through binarization, making them suitable for resource-constrained edge hardware like FPGAs and microcontrollers. While combining binarization with pruning offers further efficiency, existing pruning techniques are not well-suited for binarized models and often don't translate into real hardware savings. A new PyTorch-based framework has been developed to address this, incorporating specific freezing and pruning mechanisms for binarized neural networks. This framework allows for rapid evaluation of current methods and quick prototyping of new ones. Leveraging this framework, a novel pruning method was proposed. This method considers the relative importance of learned parameters across different abstraction levels using a global weighting mechanism. This approach consistently delivers a superior balance between model accuracy and pruning rate, achieving a 70% pruning rate on VGG11 with no loss in accuracy, significantly outperforming prior methods that reached only 41% in binarized settings.