Pruning CNN Feature Maps with Multi-Armed Bandits Boosts Efficiency

Convolutional Neural Networks (CNNs) often contain redundant feature maps, leading to increased storage requirements and higher inference costs. A new research paper proposes a loss-aware feature-map pruning framework that leverages multi-armed bandits to address this issue. This structured pruning approach removes entire convolutional output channels and their associated filters, rather than individual weights. The methodology treats each candidate feature map as an "arm" in a multi-armed bandit problem. During training, a map is temporarily masked, and its impact on the loss function is evaluated on a mini-batch. This observed loss change is then converted into a "safe-removal reward." After a predefined budget of "plays," feature maps are ranked by their learned scores, and the top-k maps are permanently removed, along with their corresponding filters, biases, and next-layer input-channel kernels. The study evaluated UCB1 and Thompson Sampling algorithms, demonstrating that both methods effectively preserve accuracy close to unpruned models while substantially reducing feature maps and convolutional computation. These bandit-based approaches significantly outperformed greedy and magnitude-based pruning techniques across various datasets, proving their effectiveness in optimizing CNNs for efficiency without sacrificing performance.