Research map / Detection and segmentation / CNN features
Reinforcement Learning as a Parsimonious Alternative to Prediction Cascades: A Case Study on Image Segmentation
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 PaSeR, an RL policy that routes image patches to a set of UNet models of varying cost (instead of IDK cascades), rewarded by IoU gain minus parameter-based cost, evaluated on battery phase segmentation and noisy MNIST.
Problem
Cascaded decision pipelines (IDK cascades) for low-cost inference waste computation through sequential intermediate queries; need a cost-aware way to select models for inputs in edge-fog-cloud settings.
Approach
Replace sequential cascading with an RL policy that, conditioned on a small model's predictions and MC-dropout entropy, samples a model choice for each image patch in one shot, rewarded by IoU improvement minus a parameter-count cost; the policy and large models are jointly fine-tuned.
Key results
- On battery segmentation, PaSeR reaches IoU 0.7426 at 1.51e11 FLOPs (IoU/GigaFlop 4.91e-3), versus 0.6987 IoU for the IDK cascade and 0.8144 for MatPhase at 2.11e12 FLOPs.
- Versus SOTA models, PaSeR reports a minimum 174% IoU/GigaFlop improvement on battery data, while staying within about 4% IoU of DeepLabV3+, SegFormer, and EfficientViT.
- On Noisy MNIST, PaSeR achieves IoU 0.8231 with IoU/GigaFlop 1.27e-4, a minimum 13.4% improvement over the best SOTA baseline on that metric.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the learned RL policy instead of the IDK cascade | PaSeR on battery material phase segmentation · Battery material phase segmentation | IoU/GigaFlop: +196% | Helped |
| use PaSeR instead of the IDK Cascade IoU-match | PaSeR · Battery material phase segmentation | IoU/GigaFlop: +923% | Helped |
| use PaSeR instead of MatPhase | Battery material phase segmentation · Battery material phase segmentation | IoU: 9.7% | Hurt |
| use PaSeR instead of MatPhase | Battery material phase segmentation · Battery material phase segmentation | IoU/GigaFlop: minimum improvement of 174% | Helped |
| replace the learned RL policy with a random policy (ablation) | PaSeR · Battery material phase segmentation | IoU: 0.7426 -> 0.7234 | Hurt |
| replace the learned RL policy with a random policy (ablation) | PaSeR · Noisy MNIST | IoU: 0.8231 -> 0.6376 | Hurt |
| replace the learned RL policy with a random policy (ablation) | PaSeR on noisy battery data · Battery material phase segmentation, noisy test data | IoU: 0.7322 vs 0.5864; degradation -1.4% vs -18.94% | Hurt |
| increase the cost weight λ (ablation) | PaSeR · Battery material phase segmentation | mean cost | Helped |
| increase the cost weight λ (ablation) | PaSeR · Battery material phase segmentation | IoU | Hurt |
| use 20 instead of 5 Monte Carlo dropout samples for entropy estimation (ablation) | PaSeR · Battery material phase segmentation | entropy distribution: p-value of 0.6986 | No clear effect |
| increase the TVD threshold from 0% to 5% to 10% (ablation) | PaSeR fine-tuning · Noisy MNIST | IoU: 0.8432, 0.8328, and 0.8231 | Hurt |
| increase the TVD threshold during fine-tuning (ablation) | PaSeR · Noisy MNIST | IoU/GigaFlop | Helped |
Surprising findings
- The random-policy variant of PaSeR degrades by 18.94% on noisy battery data, while the learned policy degrades by only 1.4%.
- The learned policy's tuned IDK cascade variant matching PaSeR's IoU needs about 10x more FLOPs (PaSeR uses 90% fewer).
Limitations
- Cost is defined as parameter count share, which may not reflect actual latency or hardware cost.
- Evaluated mainly on a small proprietary battery dataset of 40 test images and on a synthetic noisy MNIST.
- Relies on a hand-set cost weight lambda and TVD thresholds tuned by grid search.
Experimental setup
Datasets: battery material phase segmentation dataset (low-res microtomography) MNIST with Gaussian blur and box blur noise Cityscapes
Benchmarks: battery material phase segmentation Noisy MNIST Cityscapes
Metrics: IoU FLOPs IoU per GigaFlop
Baselines: IDK cascade random policy PaSeR MatPhase DeepLabV3+ SegFormer EfficientViT
Builds on: IDK cascade knowledge distillation REINFORCE
Closest papers by meaning
- Focus-Then-Decide: Segmentation-Assisted Reinforcement Learning (AAAI 2024 · similarity 0.85)
- SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- Plug-and-Play PPO: An Adaptive Point Prompt Optimizer Making SAM Greater (CVPR 2025 · similarity 0.84)
- LENS: Learning to Segment Anything with Unified Reinforced Reasoning (AAAI 2026 · similarity 0.84)
- Memory Efficient Transformer Adapter for Dense Predictions (ICLR 2025 · similarity 0.84)
- Auto-Prox: Training-Free Vision Transformer Architecture Search via Automatic Proxy Discovery (AAAI 2024 · similarity 0.84)
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.