Logic Shrinkage: Learned FPGA Netlist Sparsity for Efficient Neural Network Inference.