Fast Algorithm For Quantized Convolutional Neural Networks

Quantization has proven to be a powerful technique to reduce the memory footprint of Con volutional Neural Networks at inference time without sacrifying their accuracy. This is especially useful in the context of embedded and mobile devices, where memory resources comes at a price. E orts have been invested in trying to quantize both weights and activations to a binary value, given that computing convolution between binary sequences requires only bit-wise operations, but keeping acceptable levels of accuracy has proven hard. A bigger quantized representation than binary is needed, but a reduced computational footprint would also be desirable. To this goal, researchers have have expanded bit-wise kernels to non-binary quantized convolution. We explore a di erent approach to this problem based on the application of Number Theoretic Transforms as fast algorithms for the computation of convolution in quantized Convolutional Neural Networks.