Research map / Reinforcement learning / Policy gradient and actor-critic
Relaxation-Free Deep Hashing via Policy Gradient
Xin Yuan, Liangliang Ren, Jiwen Lu, Jie Zhou
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 |
|---|---|---|---|
| replace continuous relaxation (HashNet) with policy-gradient binary sampling | HashNet · CIFAR-10 (64-bit) | MAP: 73.9 -> 76.2 | Improved |
| replace relaxation with policy gradient | HashNet · NUS-WIDE (64-bit) | MAP: 77.6 -> 79.2 | Improved |
| replace relaxation with policy gradient | HashNet · ImageNet (64-bit) | MAP: 68.4 -> 71.6 | Improved |
| use weighted reward beta=0.5 instead of beta=0.1 (ablation) | PGDH · CIFAR-10 (64-bit) | MAP: 10.12 -> 73.65 | Improved |
| use weighted reward beta=0.5 instead of beta=0.1 (ablation) | PGDH · NUS-WIDE (64-bit) | MAP: 31.32 -> 77.95 | Improved |
| use weighted reward beta=0.7 instead of beta=0.1 (ablation) | PGDH · ImageNet (64-bit) | MAP: 1.14 -> 70.32 | Improved |
| use beta=0.8 instead of beta=0.9 (ablation) | PGDH · CIFAR-10 (64-bit) | MAP: 34.12 -> 75.12 | Improved |
| increase Monte Carlo samples T (ablation) | PGDH · CIFAR-10 | MAP: MAP degrades at small T and saturates for larger T | Mixed |
| update codebook every iteration (R=1) instead of R=5 (ablation) | PGDH · CIFAR-10 (64-bit) | MAP: MAP very low during training | Hurt |
| update codebook once per epoch (R=40) instead of R=5 (ablation) | PGDH · CIFAR-10 (64-bit) | MAP: MAP rises very slowly | Hurt |
| stochastic instead of deterministic code generation at test (ablation) | PGDH · CIFAR-10 (64-bit) | MAP: 75.54 -> 75.21 | No clear effect |
| stochastic instead of deterministic generation at epoch 1 (ablation) | PGDH · CIFAR-10 (64-bit) | MAP: 24.51 -> 10.10 | Hurt |
Closest papers by meaning
- Greedy Hash: Towards Fast Optimization for Accurate Hash Coding in CNN (NeurIPS 2018 · similarity 0.92)
- Accelerate Learning of Deep Hashing With Gradient Attention (ICCV 2019 · similarity 0.89)
- Learning to Hash with Binary Deep Neural Network (ECCV 2016 · similarity 0.88)
- Deep Hashing: A Joint Approach for Image Signature Learning (AAAI 2017 · similarity 0.88)
- Hashing with Binary Matrix Pursuit (ECCV 2018 · similarity 0.88)
- Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion (AAAI 2024 · similarity 0.87)
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.
