Semi-Supervised Learning under Class Distribution Mismatch
Authors: Yanbei Chen, Xiatian Zhu, Wei Li, Shaogang Gong3569-3576
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide extensive benchmarking results in this realistic SSL scenario, including our proposed UASD and six representative state-of-the-art SSL methods on three image classification datasets: CIFAR10, CIFAR100 and Tiny Image Net. Remarkably, UASD outperforms all the strong competitors often by large margins, and demonstrates great potential to exploit the unconstrained unlabelled data. |
| Researcher Affiliation | Collaboration | Yanbei Chen,1 Xiatian Zhu,2 Wei Li,1 Shaogang Gong1 1Queen Mary University of London, 2Vision Semantics Ltd. |
| Pseudocode | Yes | Algorithm 1 Uncertainty-Aware Self-Distillation (UASD) Require: Labelled data Dl = {xi,l, yi}Nl i=1. Unlabelled data Du = {xi,u}Nu i=1. Require: Trainable neural network θ. Ramp-up weighting function w(t). for t = 1 to max epoch do Refresh confidence threshold τt per epoch. for k = 1 to max iter per epoch do Forward propagation to accumulate network prediction qt(y|xi) (Eq (1)) for every in-batch sample. Apply OOD filtering (Eq (2), (3)). Update network parameters θ with loss function Eq (4). end for end for |
| Open Source Code | No | The paper states, 'For a comprehensive and fair comparison, our experiments are built upon the open-source Tensorflow implementation by Oliver et al. (Oliver et al. 2018).' This refers to a base implementation they used, not an explicit statement that their own UASD code is open source or provided. |
| Open Datasets | Yes | Datasets. We use three image classification benchmark datasets. (1) CIFAR10: A natural image dataset with 50,000/10,000 training/test samples from 10 object classes. (2) CIFAR100: A dataset of 100 fine-grained classes, with the same amount of training/test samples as CIFAR10. (3) Tiny Image Net: A subset of Image Net (Deng et al. 2009) with 200 classes, each of which has 500/50 training/validation images. |
| Dataset Splits | Yes | Thus, we dynamically estimate τt in a data-driven manner by using the validation set (10% of training data) of known classes as reference. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models, memory specifications, or cloud instance types used for running experiments. |
| Software Dependencies | No | The paper mentions building upon 'the open-source Tensorflow implementation by Oliver et al. (Oliver et al. 2018)' but does not specify a version number for TensorFlow or any other software dependencies. |
| Experiment Setup | Yes | Implementation details. For a comprehensive and fair comparison, our experiments are built upon the open-source Tensorflow implementation by Oliver et al. (Oliver et al. 2018). It uses the standard Wide Res Net (Zagoruyko and Komodakis 2016), i.e. WRN-28-2, as the base network and Adam optimiser (Kingma and Ba 2014) for training. We revise the default 10-dimensional classification layer to K-dimension, where K is the number of known classes in the labelled data. Unless stated otherwise, all hyper-parameters, the ramp-up function, and training procedures are the same as that of (Oliver et al. 2018). |