Principal Filter Analysis for Guided Network Compression

Principal Filter Analysis (PFA), is an elegant, easy to implement, yet effective methodology for neural network compression. PFA exploits the intrinsic correlation between filter responses within network layers to recommend a smaller network footprint. We propose two compression algorithms: the first allows a user to specify the proportion of the original spectral energy that should be preserved in each layer after compression, while the second is a parameter-free approach that automatically selects the compression used at each layer. Both algorithms are evaluated against several architectures and datasets, and we show considerable compression rates without compromising accuracy, e.g., for VGG-16 on CIFAR-10 and CIFAR-100 PFA achieves a compression rate of 8x and 3x with an accuracy gain of 0.4% points and 1.4% points, respectively. In our tests we also demonstrate that networks compressed with PFA achieve an accuracy that is very close to the empirical upper bound for a given compression ratio.