NullaNet: Training Deep Neural Networks for Reduced-Memory-Access Inference
This addresses the challenge of deploying deep neural networks on resource-constrained devices by enabling more efficient inference, though it appears incremental as it builds on existing training methods with a novel realization approach.
The paper tackled the problem of high computational and storage complexity in deep neural networks by introducing a training method that enables realization through Boolean logic minimization, resulting in a reduction of memory access energy, about two orders of magnitude fewer computing resources, and substantially lower latency.
Deep neural networks have been successfully deployed in a wide variety of applications including computer vision and speech recognition. However, computational and storage complexity of these models has forced the majority of computations to be performed on high-end computing platforms or on the cloud. To cope with computational and storage complexity of these models, this paper presents a training method that enables a radically different approach for realization of deep neural networks through Boolean logic minimization. The aforementioned realization completely removes the energy-hungry step of accessing memory for obtaining model parameters, consumes about two orders of magnitude fewer computing resources compared to realizations that use floatingpoint operations, and has a substantially lower latency.