Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Discrete Representations via Information Maximizing Self-Augmented Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 deep neural networks regularized with VAT-based SAT rather than a linear classifier | IMSAT clustering · MNIST, Omniglot, STL, CIFAR100, SVHN | ACC (%): 59.6 (2.3) -> 98.4 (0.4); 11.1 (0.2) -> 24.0 (0.9); 73.5 (6.5) -> 94.1 (0.4); 23.7 (0.8) -> 27.5 (0.4); 20.2 (1.4) -> 57 | Helped |
| Add affine augmentation to VAT (ablation) | IMSAT clustering · Omniglot | ACC (%): 24.0 (0.9) -> 70.0 (2.0) | Helped |
| Use affine augmentation without VAT (ablation) | IMSAT clustering · Omniglot | ACC (%): 24.0 (0.9) -> 45.1 (2.0) | Helped |
| Use VAT instead of random perturbation training (ablation) | IMSAT clustering · MNIST, Omniglot, STL, CIFAR10, CIFAR100, SVHN, 20news | ACC (%): 89.6 (5.4) -> 98.4 (0.4); 16.4 (3.1) -> 24.0 (0.9); 92.8 (2.5) -> 94.1 (0.4); 45.5 (2.9) -> 45.6 (0.8); 24.7 (0.5) -> 27 | Helped |
| Use a linear rather than deep classifier with VAT-based SAT (ablation) | IMSAT clustering · MNIST, STL, CIFAR10, SVHN, Reuters | ACC (%): 61.1 (1.9) vs 98.4 (0.4); 91.7 (0.5) vs 94.1 (0.4); 40.7 (0.6) vs 45.6 (0.8); 18.2 (1.9) vs 57.3 (3.9); 42.9 (0.8) vs 71 | Hurt |
| Increase hash-network hidden layers from 60-30 to 200-200 (ablation) | IMSAT (VAT), 16-bit hash · MNIST | mAP (%): 61.2 (2.5) -> 80.7 (2.2) | Helped |
| Increase hash-network hidden layers from 200-200 to 400-400 (ablation) | IMSAT (VAT), 16-bit hash · MNIST | mAP (%): 80.7 (2.2) -> 83.9 (2.3) | Helped |
| Increase hash-network hidden layers from 60-30 to 200-200 (ablation) | IMSAT (VAT), 16-bit hash · CIFAR10 | mAP (%): 19.8 (1.2) -> 21.2 (0.8) | Helped |
| Increase hash-network hidden layers from 200-200 to 400-400 (ablation) | IMSAT (VAT), 16-bit hash · CIFAR10 | mAP (%): 21.2 (0.8) -> 21.4 (0.5) | No clear effect |
| Increase Deep RIM hash-network size (ablation) | Deep RIM · MNIST and CIFAR10 hash learning | mAP (%): MNIST: 42.7 (2.8) -> 43.9 (2.7); CIFAR10: 15.2 (0.5) -> 15.4 (0.2) | No clear effect |
| Use IMSAT (VAT) rather than Deep RIM for 16-bit hashing | Hash learning · MNIST | mAP (%): 43.9 (2.7) -> 83.9 (2.3) | Helped |
| Use IMSAT (VAT) rather than the strongest listed baseline | 16-bit hash learning · CIFAR10 | mAP (%): 24.0 (3.5) -> 21.4 (0.5) | Hurt |
Closest papers by meaning
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.85)
- Understanding Augmentation-based Self-Supervised Representation Learning via RKHS Approximation and Regression (ICLR 2024 · similarity 0.84)
- Self-supervised Representation Learning from Random Data Projectors (ICLR 2024 · similarity 0.84)
- Efficient Distribution Matching of Representations via Noise-Injected Deep InfoMax (ICLR 2025 · similarity 0.83)
- Representation Space Augmentation for Effective Self-Supervised Learning on Tabular Data (AAAI 2025 · similarity 0.83)
- Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections (NeurIPS 2025 · similarity 0.82)
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.