Pruning Binarized Neural Networks Enables Low-Latency, Low-Power FPGA-Based Handwritten Digit Classification.