Convolutional neural networks often contain redundant feature maps that increase storage and inference cost. This paper presents a loss-aware feature-map pruning framework using multi-armed bandits. Feature-map pruning is structured as it removes complete convolutional output channels and their producing filters rather than isolated scalar weights.
In this framework, each candidate feature map is treated as an "arm". At each play time, one map is temporarily masked and evaluated on a sampled mini-batch; the map is then restored, and the observed loss change is converted into a safe-removal reward. After a fixed play budget, candidate maps are ranked by learned scores, and the top-k maps are permanently removed along with their filters, biases, and corresponding next-layer input-channel kernels.
The study evaluates UCB1 and Thompson Sampling, comparing them with direct/oracle-style evaluation on LeNet/MNIST, and extends the evaluation to MNIST, CIFAR-10, CIFAR-100, SVHN, CUB-200-2011, and Oxford Flowers 102. Results show that UCB1 and Thompson Sampling preserve accuracy close to unpruned models while removing feature maps and reducing convolutional computation. Friedman and Nemenyi tests indicate that UCB1 obtains the highest mean rank, followed by Thompson Sampling; both significantly outperform greedy and magnitude-based pruning while remaining statistically comparable to the original unpruned model.
Blogger's Review: This paper explores a novel approach to feature map pruning through multi-armed bandits, demonstrating the potential to effectively reduce computational resources while maintaining model performance. The loss-aware approach offers new insights for future network optimization and warrants further research across broader applications.