Loss-Aware Feature-Map Pruning Using Multi-Armed Bandits
A recent study published on arXiv introduces a framework for pruning feature maps in convolutional neural networks that is aware of loss, utilizing multi-armed bandits. In this approach, each candidate feature map is considered as an arm, which is temporarily masked to assess the change in loss over a mini-batch, translating this change into a reward for safe removal. After a predetermined budget is reached, the top-k feature maps, along with their associated filters, biases, and input-channel kernels for the next layer, are permanently discarded. The research analyzes UCB1 and Thompson Sampling, contrasting them with direct/oracle-style evaluations on LeNet/MNIST, and further extends the evaluation to include MNIST, CIFAR-10, and CIFAR-100.
Key facts
- Paper on arXiv:2607.22564v1
- Loss-aware feature-map pruning using multi-armed bandits
- Each candidate feature map treated as an arm
- Map temporarily masked and evaluated on sampled mini-batch
- Loss change converted into safe-removal reward
- Top-k maps permanently removed after fixed play budget
- Evaluates UCB1 and Thompson Sampling
- Compared with direct/oracle-style evaluation on LeNet/MNIST
- Extended evaluation to MNIST, CIFAR-10, CIFAR-100
Entities
Institutions
- arXiv