Differentiable Training for Hardware Efficient LightNNs
To reduce runtime and resource utilization of Deep Neural Networks (DNNs) on customized hardware, LightNN has been proposed by constraining the weights of DNNs to be a sum of a limited number (denoted as $k\in\{1,2\}$) of powers of 2. LightNNs can therefore can replace the multiplication between activations and weights with a shift operation or two shifts and an add operation. To provide a more continuous Pareto-optimal curve of accuracy and runtime so that hardware designers can have more flexible options of DNN configurations, one can customize the $k$ for each convolutional filter. In this paper, we formulate the selection of $k$ to be differentiable, and train the model weights and per-filter $k$ in an end-to-end fashion. Since flexible-$k$ LightNNs (FLightNNs) fully utilize the hardware resources on Field Programmable Gate Arrays (FPGAs), our experimental results show that FLightNN can achieve 2$\times$ speedup on FPGAs when compared to LightNN-2, with only 0.1\% accuracy degradation. In addition, compared to a 4-bit fixed-point quantization, FLightNN can achieve slightly higher accuracy and 1.4$\times$ speedup, due to its lightweight shift operations.
PDF Abstract