TENT: Efficient Quantization of Neural Networks on the tiny Edge with Tapered FixEd PoiNT

In this research, we propose a new low-precision framework, TENT, to leverage\nthe benefits of a tapered fixed-point numerical format in TinyML models. We\nintroduce a tapered fixed-point quantization algorithm that matches the\nnumerical format's dynamic range and distribution to that of the deep neural\nnetwork model's parameter distribution at each layer. An accelerator\narchitecture for the tapered fixed-point with TENT framework is proposed.\nResults show that the accuracy on classification tasks improves up to ~31 %\nwith an energy overhead of ~17-30 % as compared to fixed-point, for ConvNet and\nResNet-18 models.\n