Inducing and Exploiting Activation Sparsity for Fast Inference on Deep Neural Networks