IIT Madras

Title - Weightless Neural Networks : Towards ending the wait for edge hardware-centric Machine Learning
Abstract - Deploying fast, accurate, and efficient machine learning on edge devices remains a key research challenge. Deep Neural Networks (DNNs) have been traditionally popular, with wide ranging applications across domains. However, their computational and storage demands often hinder their deployment on edge devices that have stringent resource, latency, and energy budgets; with model quantization, pruning, and other such techniques being active areas of research to mitigate this. Weightless Neural Networks (WNNs), an unconventional class of look-up table based neural networks, offer an efficient alternative to DNNs. These networks eliminate most of the conventional model “weights”, and their architecture closely resembles the underlying logic fabric of hardware devices like FPGAs. However, these have been historically underutilized, and have been limited to relatively tiny models and simple applications. In this talk, I will present some of our recent research in the field of WNNs and their associated hardware accelerators, with an emphasis on our latest work (LL-ViT) that further extends these to the realm of vision transformers — making a case for mainstream adoption of these hardware-centric neural networks. Compared to their DNN (ViT) counterparts, LL-ViTs offer upto 2x energy efficiency on FPGAs/ASICs, with about 60% reduction in model weights and computations, while achieving better accuracies on image classification tasks across datasets like CIFAR-10/100, TinyImageNet and others. Towards the end, I will also briefly touch upon architectural enhancements to mainstream compute devices, like GPUs, to further improve the benefits offered by these networks.