Computer Vision News - March 2020

3 Summary Automatic Pruning fo ... 5 a Bayesian optimization scheme to determine the optimal pruning ratio. Let's dive into the details. The problem formulation is as follows: we let F k denote the k'th filter convolution on the l-th layer of a CNN. Then, if we denote by Ok the k'th output feature map, and by I the input feature map, it holds that: Method In the quantization process, after the training phase, each filter is quantized with a predetermined pointwise quantization function Q(F k ). In order to prune the filters, the method first rank each quantized filter with regards to its original form. Then, the set of filters is chosen by setting a threshold th using the rule: { F k |distance( F k ,Q( F k ))<th} . To define the measure for distance, the authors use two kind of metrics (defined on a vector v): = ∗ Φ = ⋅ ( ) || |||| ( )|| , = || − ( )|| As an initial experiment in the graph below, the authors demonstrate the effectiveness of this pruning scheme. They tested the consequences of pruning the network in ascending/descending order, depending on the angle metric and the distance metric: