Machine learning engineer based in London, UK.
2017-07-05 15:27:00 +0000
Best self-reported performances on semi-supervised image classification benchmarks. Inspired by Rodrigo Benenson’s mini-site collating published performance on key computer vision datasets.
Semi-supervised learning corresponds to the case where a large dataset may be available, but only a small subset of the data is labelled. Traditional semi-supervised learning includes graph-based algorithms like label propagation, which is included in scikit-learn. Deep learning approaches tend to make use of novel architectures (e.g. ladder networks), regularization techniques (e.g. virtual adversarial training), or generative models (e.g. DGN, GAN).
| 100 labels | 1000 labels | All labels | Method | Year |
|---|---|---|---|---|
| 0.93 (±0.065) | N/A | N/A | Improved Techniques for Training GANs | 2016 |
| 1.002 (±0.038) | 0.979 (±0.025) | 0.578 (±0.013) | Deconstructing the Ladder Network Architecture (Ladder w/ AMLP[2,2,2]) | 2015 |
| 1.072 (±0.015) | 0.974 (±0.021) | 0.598 (±0.014) | Deconstructing the Ladder Network Architecture (Ladder w/ AMLP[4]) | 2015 |
| 1.072 (±0.015) | 1.193 (±0.039) | 0.569 (±0.010) | Deconstructing the Ladder Network Architecture (Ladder w/ AMLP[2,2]) | 2015 |
| 1.06 (±0.37) | 0.84 (±0.08) | 0.57 (±0.02) | Semi-Supervised Learning with Ladder Networks | 2015 |
| 1.36 | 1.27 | 0.64 | Virtual Adversarial Training: a Regularization Method for Supervised and Semi-supervised Learning | 2017 |
| 2.33 | 1.36 | 0.637 (±0.046) | Distributional Smoothing with Virtual Adversarial Training | 2016 (ICLR) |
work in progress: incomplete
| 500 labels | 1000 labels | All labels | Method | Year |
|---|---|---|---|---|
| N/A | 3.86 | N/A | Virtual Adversarial Training: a Regularization Method for Supervised and Semi-supervised Learning [Conv-Large w/ EntMin, w/ augmentation] | 2017 |
| N/A | 4.28 | N/A | Virtual Adversarial Training: a Regularization Method for Supervised and Semi-supervised Learning [Conv-Large w/ EntMin, no augmentation] | 2017 |
| 5.12 (±0.13) | 4.42 (±0.16) | 2.74 (±0.06) | Temporal Ensembling for Semi-Supervised Learning [w/ augmentation] | 2016 |
| 6.65 (±0.53) | 4.82 (±0.17) | 2.54 (±0.04) | Temporal Ensembling for Semi-Supervised Learning [Pi model w/ augmentation] | 2016 |
| N/A | 24.63 | N/A | Distributional Smoothing with Virtual Adversarial Training | 2016 (ICLR) |
work in progress: incomplete
| 4k labels | All labels | Method | Year |
|---|---|---|---|
| 10.55 | N/A | Virtual Adversarial Training: a Regularization Method for Supervised and Semi-supervised Learning [Conv-Large w/ EntMin, w/ augmentation] | 2017 |
| 12.16 (±0.24) | 5.60 (±0.10) | Temporal Ensembling for Semi-Supervised Learning [w/ augmentation] | 2016 |
| 13.15 | N/A | Virtual Adversarial Training: a Regularization Method for Supervised and Semi-supervised Learning [Conv-Large w/ EntMin, no augmentation] | 2017 |
| 20.40 | N/A | Semi-Supervised Learning with Ladder Networks [Conv-Large, Gamma model, no augmentation] | 2015 |
| 10k labels | All labels | Random 500k Tiny Images | Restricted 237k Tiny Images | Method | Year |
|---|---|---|---|---|---|
| 38.65 (±0.51) | 26.30 (±0.15) | 23.62 (±0.23) | 23.79 (±0.24) | Temporal Ensembling for Semi-Supervised Learning | 2016 |