Structured Neuron Pruning in Deep Neural Networks Using Multi-Armed Bandits

2026-05-29 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new structured pruning framework utilizes multi-armed bandit (MAB) algorithms to remove complete neurons from deep neural networks, addressing the challenge of exploiting unstructured sparsity in dense implementations. This method treats each candidate neuron as an "arm," temporarily masking it to measure loss change on a mini-batch, restoring it, and updating a "safe-removal reward" estimate. The framework supports various stochastic policies, including Epsilon-Greedy, Softmax, UCB1, and Thompson Sampling, alongside multiplicative-weight policies like Hedge-style multiplicative weights and EXP3. Evaluations across tabular classification, tabular regression, and deep neural-network benchmarks (image, text, reasoning tasks) demonstrated its efficacy. Statistical analysis, using the Friedman test and Nemenyi post-hoc test, revealed UCB1 as the top performer in tabular tasks, improving on unpruned networks and competing with standard regression models by R^2. For deep-learning tasks, UCB1 and Thompson Sampling achieved the strongest ranks, significantly outperforming unpruned models and other pruning techniques.

Key takeaway

For Machine Learning Engineers optimizing deep neural network deployment, consider integrating multi-armed bandit (MAB) based neuron pruning. This method, particularly using UCB1 or Thompson Sampling, offers a computationally practical way to achieve structured model reduction, outperforming traditional pruning techniques and unpruned models across various tasks. You can significantly reduce model size and improve efficiency without sacrificing performance, especially for image, text, and reasoning applications.

Key insights

Multi-armed bandits effectively prune entire neurons in deep neural networks, improving model efficiency and performance.

Principles

• Redundant hidden units exist in deep neural networks.
• Structured pruning is more exploitable than unstructured sparsity.
• MAB algorithms can estimate neuron removal rewards efficiently.

Method

Treat each neuron as an MAB arm; pull to mask, measure loss change on a mini-batch, restore, and update a safe-removal reward estimate using policies like UCB1 or Thompson Sampling.

In practice

• Apply UCB1 for tabular classification and regression tasks.
• Use UCB1 or Thompson Sampling for deep learning model pruning.
• Reduce model size for image, text, and reasoning applications.

Topics

• Deep Neural Networks
• Neuron Pruning
• Multi-Armed Bandits
• Model Reduction
• UCB1 Algorithm
• Thompson Sampling

Best for: Research Scientist, NLP Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• Replicable Bandits with UCB based Exploration — New UCB-based algorithms improve replicability and regret bounds for stochastic multi-armed and linear bandits.
• Multiple Additive Neural Networks for Structured and Unstructured Data — MANN enhances gradient boosting with shallow neural networks for diverse data types, improving accuracy and robustness.
• Batched Single-Index Global Multi-Armed Bandits with Covariates — BIDS is a novel batched multi-armed bandit algorithm that uses single-index regression to efficiently manage covariates and related arm rewards.
• Calibration-Gated LLM Pseudo-Observations for Online Contextual Bandits — LLM pseudo-observations can reduce cold-start regret in contextual bandits, but prompt design is critical.
• MINTS: Minimalist Thompson Sampling — MINTS simplifies Bayesian decision-making by focusing priors and profiling nuisance parameters to handle structural constraints effectively.
• A single algorithm for both restless and rested rotting bandits — RAW-UCB unifies restless and rested rotting bandit problems, achieving near-optimal regret without prior setting knowledge.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.