Abstract
We propose Deterministic Binary Filters, an approach to Convolutional Neural Networks that learns weighting coefficients of predefined orthogonal binary basis instead of the conventional approach of learning directly the convolutional filters. This approach results in architectures offering significantly fewer parameters (4× to 16×) and smaller model sizes (up to 32× due to the use of binary rather than floating point precision). We show our deterministic filter design can be integrated into well-known network architectures (such as ResNet and SqueezeNet) with as little as 2% loss of accuracy under datasets like CIFAR-10. Under ImageNet, they are used in an architectures 3× smaller compared to sub-megabyte binary networks while reaching comparable accuracy levels.