Loss-Aware Feature-Map Pruning in Convolutional Neural Networks Using Multi-Armed Bandits
Quick Answer
This paper introduces a loss-aware feature-map pruning framework for convolutional neural networks using multi-armed bandits, specifically UCB1 and Thompson Sampling.
Quick Take
The approach effectively reduces computation while maintaining accuracy across various datasets, including MNIST and CIFAR-10, outperforming traditional pruning methods. Results indicate that UCB1 achieves the highest mean rank and significantly enhances model efficiency without sacrificing performance.
Key Points
- Utilizes multi-armed bandits for structured feature-map pruning in CNNs.
- Evaluates UCB1 and Thompson Sampling against various datasets like MNIST and CIFAR-10.
- UCB1 and Thompson Sampling maintain accuracy while reducing convolutional computation.
- Both methods significantly outperform greedy and magnitude-based pruning techniques.
- Friedman and Nemenyi tests confirm UCB1's superior performance ranking.
DeepSignal Analysis
What happened
The paper presents a framework for pruning feature maps in convolutional neural networks using multi-armed bandits, specifically UCB1 and Thompson Sampling. This method aims to reduce computational costs while maintaining accuracy across various datasets, including MNIST and CIFAR-10. UCB1 reportedly achieves the highest mean rank in terms of efficiency.
Key evidence
- The study evaluates the performance of UCB1 and Thompson Sampling against direct evaluations on LeNet/MNIST, demonstrating their effectiveness in preserving accuracy while reducing computation.
- Results indicate that UCB1 and Thompson Sampling significantly outperform traditional greedy and magnitude-based pruning methods, while remaining statistically comparable to unpruned models.
- The framework ranks candidate feature maps based on learned scores after a fixed budget of evaluations, allowing for the permanent removal of the least effective maps.
Why it matters
This research addresses the challenge of redundancy in convolutional neural networks, which can lead to increased storage and inference costs. By employing a loss-aware approach to feature-map pruning, the framework not only enhances computational efficiency but also maintains model performance. This could have implications for deploying neural networks in resource-constrained environments, where efficiency is critical.
Paper Resources
📖 Reader Mode
~2 min readAbstract: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 because it removes complete convolutional output channels and their producing filters rather than isolated scalar weights. 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 with their filters, biases and corresponding next-layer input-channel kernels. The study evaluates UCB1 and Thompson Sampling, compares 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 show 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.
| Comments: | 18 pages, 4 figures |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.22564 [cs.AI] |
| (or arXiv:2607.22564v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2607.22564 arXiv-issued DOI via DataCite |
Submission history
From: Salem Ameen [view email]
[v1]
Fri, 29 May 2026 20:03:08 UTC (560 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →HOBA: Hierarchical On-Policy Bidding Agents for Adaptive Online Advertising
HOBA (Hierarchical On-policy Bidding Agents) is a novel hierarchical reinforcement learning framework that enhances online advertising bidding systems by improving adaptability and reducing hyperparameter tuning costs. It utilizes a for hyperparameter inference, a SARSA agent for expert model selection, and a dynamic expert pool for bid execution, achieving a +3.6% increase in target cost during large-scale deployment and outperforming state-of-the-art baselines on AuctionNet.