Computer Vision News - March 2020

3 Summary Automatic Pruning fo ... 7 On ImageNet, using the angle metric, with nearly the same price of accuracy, the method achieved less pruning ratio of up to 20%. This can be seen in the table below for a base model of ResNet-14: These are very nice results, since we know that ImageNet is a complex dataset and usually the trained weights on this dataset tend to be less redundant. Moreover, it is quite remarkable that we can achieve 59% on top-1 accuracy with a model that weights 4.36 MB. The paper proposes a method to effectively pruning redundant filters with low- precision. It does so by introducing two prunning metrics that seems to be highly correlated with the model accuracy. Additionally, the paper suggests a Bayesian optimization scheme to efficiently determine the pruning ratio for each layer. The method performs very well on ImageNet and CIFAR-10, presenting pruning ratio of up to 53% with a small price on accuracy. There is no doubt that the paper tackles a fundamental deep learning problem for the whole field that has to be solved in order to move forward. Conclusion