Research map / Label-efficient and robust learning / Recognition and long-tail
Drop-Bottleneck: Learning Discrete Compressed Representation for Noise-Robust Exploration
Jaekyeom Kim, Minjung Kim, Dongyeon Woo, Gunhee Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Drop-Bottleneck rather than VIB in the exploration method | PPO + Ours · VizDoom MyWayHome sparse and very sparse tasks under noise | average episodic sum of rewards | Improved |
| Use Drop-Bottleneck rather than VIB in episodic exploration | PPO + Ours · DMLab, all six Sparse and Very Sparse noise settings | average episodic sum of rewards: 28.8 vs 23.5; 29.1 vs 25.4; 26.9 vs 22.3 | Improved |
| Add Drop-Bottleneck feature dropping (ablation) | PPO + Ours exploration · DMLab Very Sparse, Image Action | average episodic sum of rewards: 14.9 ± 1.17 -> 28.8 ± 2.1 | Improved |
| Add Drop-Bottleneck feature dropping (ablation) | PPO + Ours exploration · DMLab Very Sparse, Noise | average episodic sum of rewards: 11.7 ± 1.3 -> 29.1 ± 2.0 | Improved |
| Add Drop-Bottleneck feature dropping (ablation) | PPO + Ours exploration · DMLab Very Sparse, Noise Action | average episodic sum of rewards: 10.3 ± 1.3 -> 26.9 ± 1.7 | Improved |
| Add Drop-Bottleneck feature dropping (ablation) | PPO + Ours exploration · DMLab Very Sparse, Original | average episodic sum of rewards: 33.0 ± 1.8 -> 42.7 ± 2.3 | Improved |
| Use DB with β=0.01 instead of VIB | Inception-ResNet-v2 · ImageNet targeted ℓ2 attack | attack success rate (%): 100.0 -> 18.5 | Improved |
| Use DB with β=0.003162 instead of VIB | Inception-ResNet-v2 · ImageNet targeted ℓ∞ attack | attack success rate (%): 97.5 -> 1.5 | Improved |
| Use deterministic DB feature selection instead of univariate MI feature selection | Inception-ResNet-v2 · ImageNet targeted ℓ2 attack with 8 features | attack success rate (%): 99.5 -> 20.0 | Improved |
| Use deterministic DB feature selection instead of univariate MI feature selection | Inception-ResNet-v2 · ImageNet targeted ℓ∞ attack with 8 features | attack success rate (%): 99.5 -> 2.0 | Improved |
| Increase DB compression strength β | DB image classifier · ImageNet validation | classification accuracy | Mixed |
| Use the deterministic DB representation instead of its stochastic representation | Drop-Bottleneck · ImageNet validation | classification accuracy | No clear effect |
Closest papers by meaning
- Curiosity-Bottleneck: Exploration By Distilling Task-Specific Novelty (ICML 2019 · similarity 0.89)
- Information Bottleneck Analysis of Deep Neural Networks via Lossy Compression (ICLR 2024 · similarity 0.86)
- Structured IB: Improving Information Bottleneck with Structured Feature Learning (AAAI 2025 · similarity 0.84)
- Cauchy-Schwarz Divergence Information Bottleneck for Regression (ICLR 2024 · similarity 0.84)
- Latent Representation and Simulation of Markov Processes via Time-Lagged Information Bottleneck (ICLR 2024 · similarity 0.84)
- A Rate-Distortion View of Uncertainty Quantification (ICML 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.
