Research map / Reinforcement learning / Visual and representation RL
Jointly Training and Pruning CNNs via Learnable Agent Guidance and Alignment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RLAL, which jointly trains CNN weights and structurally prunes channels using a SAC reinforcement-learning agent whose non-stationary reward is modeled by a GRU over learnable epoch embeddings, plus a group-lasso alignment loss.
Problem
Structural pruning of CNNs usually needs a pretrained model before pruning, and standard episodic RL pruning methods assume a stationary reward, which breaks when weights change during joint training and pruning.
Approach
An RL agent sets per-layer pruning ratios; a GRU over learnable epoch embeddings, trained with a decoder to reconstruct observed rewards, provides an environment representation that is appended to the SAC agent's state and critics, while a group-lasso term softly pushes weights toward the agent's selected sub-network.
Key results
- On CIFAR-10 ResNet-56 at 50% FLOPs pruning, RLAL reaches 93.86% accuracy (+0.45% vs. baseline 93.41%), outperforming DCP-Adapt and AMC.
- On ImageNet ResNet-34 at 50% FLOPs pruning, RLAL achieves -0.14% top-1 change versus -0.45% for ISP at 44% FLOPs.
- On ImageNet MobileNet-V2 at 29.4% FLOPs, RLAL gets -0.50% top-1 versus -1.00% for AMC at 30% FLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RLAL instead of DCP-Adapt | ResNet-56 · CIFAR-10 | ∆-Acc: +0.01% -> + 0.45% | Helped |
| Use RLAL instead of DDNP | MobileNet-V2 · CIFAR-10 | ∆-Acc: + 0.23% -> + 0.37% | Helped |
| Use RLAL instead of GNNRL | ResNet-18 · ImageNet | ∆-Acc Top-1: −1.10% -> −0.80% | Helped |
| Use RLAL instead of GP | ResNet-34 · ImageNet | ∆-Acc Top-1: −1.14% -> −0.14% | Helped |
| Use RLAL instead of MetaPruning | MobileNet-V2 · ImageNet | ∆-Acc Top-1: −0.80% -> −0.50% | Helped |
| remove epoch embeddings and recurrent environment representation (ablation) | RLAL · CIFAR-10, ResNet-56 | ∆-Acc: +0.45% -> −0.03% | Hurt |
| remove soft regularization in addition to epoch embeddings and recurrent environment representation (ablation) | RLAL · CIFAR-10, ResNet-56 | ∆-Acc: +0.45% -> −0.21% | Hurt |
| increase agent episodes per epoch from 5 to 10 or 15 (ablation) | RLAL · CIFAR-10, ResNet-56, pruning rates 35%, 50%, and 65% | best reward | Helped |
| provide epoch-embedding-based environment representations to the agent (ablation) | RLAL · CIFAR-10, ResNet-56 and MobileNet-V2 | best reward | Helped |
Surprising findings
- Removing the epoch embeddings and recurrent environment model alone drops ResNet-56 pruned accuracy from +0.45% to -0.03% relative to baseline.
Limitations
- Reward computation requires evaluating pruned sub-networks on a held-out subset, which is time-consuming per trajectory.
- Gains on MobileNet-V2 ImageNet are modest and pruning is harder for efficient architectures.
Experimental setup
Datasets: CIFAR-10 ImageNet
Benchmarks: CIFAR-10 ImageNet
Metrics: top-1 accuracy top-5 accuracy delta accuracy pruned FLOPs
Baselines: AMC DCP-Adapt SCP FPGM SFP FPC HRank DTP GNNRL GP
Builds on: Soft Actor-Critic AMC norm-based filter pruning
Closest papers by meaning
- Runtime Neural Pruning (NeurIPS 2017 · similarity 0.89)
- Storage Efficient and Dynamic Flexible Runtime Channel Pruning via Deep Reinforcement Learning (NeurIPS 2020 · similarity 0.88)
- When To Prune? A Policy Towards Early Structural Pruning (CVPR 2022 · similarity 0.88)
- Rethinking the Value of Network Pruning (ICLR 2019 · similarity 0.88)
- MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (CVPR 2024 · similarity 0.86)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.