Aggressive compression of mobilenets using hybrid ternary layers

Results: As shown in TABLE I., the hybrid filter banks achieve 46.4% and 51.07% reduction in multiplications and model size respectively while incurring modest (48%) increase in additions. This translates into 28% savings in energy required per inference while ensuring no degradation in throughput on a DNN hardware accelerator consisting of both MAC and adders when compared to the execution of baseline MobileNets on a MAC-only hardware accelerator.Significance for the tinyML community: MobileNets family of computer vision neural networks have fueled tremendous progress in the design and organization of resource-efficient architectures in recent years. New applications with stringent real-time requirements on highly constrained devices require further compression of MobileNets to make it amenable for edge devices. The hybrid filter bank is a first step towards ternarizing the already compute-efficient MobileNets with a negligible loss in accuracy on a large-scale dataset such as ImageNet, better enabling deployment for vision-based “tinyML” applications. See [4] for details.