Source-linked AI summary
Self-Supervised Learning For Few-Shot Image Classification
Da Chen, Yuefeng Chen, Yuhong Li, Feng Mao, Yuan He, Hui Xue
TL;DR
Few-shot classification must adapt to unseen classes from limited labeled examples, while supervised pretraining can constrain embedding-network scale. The paper uses self-supervised pretraining followed by episodic meta-learning, reporting stronger results on standard and cross-domain benchmarks.
Problem
Few-shot classification adapts to unseen classes with limited labeled examples, and existing supervised training can constrain the scale of embedding networks.
Source: §1 · p. 1Method
The method trains a large embedding network with self-supervised learning, then fine-tunes it through episodic meta-learning for few-shot classification.
Source: §3 · p. 2, §3.2 · p. 3Results
Across four cross-domain datasets and three tasks, the method reports nearly 1.5% improvement over state-of-the-art results, alongside gains over baselines on MiniImageNet and CUB.
Source: §4.3.2 · p. 4, §4.3.1 · p. 4Takeaways & Limitations
The findings support self-supervised embeddings as a more generalized and transferable base for meta-learning in standard and cross-domain few-shot classification.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Few-shot image classification aims to classify unseen classes with limited labelled samples. Recent works benefit from the meta-learning process with episodic tasks and can fast adapt to class from training to testing. Due to the limited number of samples for each task, the initial embedding network for meta-learning becomes an essential component and can largely affect the performance in practice. To this end, most of the existing methods highly rely on the efficient embedding network. Due to the limited labelled data, the scale of embedding network is constrained under a supervised learning(SL) manner which becomes a bottleneck of the few-shot learning methods. In this paper, we proposed to train a more generalized embedding network with self-supervised learning (SSL) which can provide robust representation for downstream tasks by learning from the data itself. We evaluate our work by extensive comparisons with previous baseline methods on two few-shot classification datasets ({\em i.e.,} MiniImageNet and CUB) and achieve better performance over baselines. Tests on four datasets in cross-domain few-shot learning classification show that the proposed method achieves state-of-the-art results and further prove the robustness of the proposed model. Our code is available at \hyperref[https://github.com/phecy/SSL-FEW-SHOT.]{https://github.com/phecy/SSL-FEW-SHOT.}
1. INTRODUCTION
Few-shot learning seeks adaptation to new tasks with few instances, but fine-tuning risks overfitting and meta-learning depends on an effective pretrained embedding.
1. INTRODUCTION
- Few-shot learning studies how to adapt models to new data or tasks using a restricted number of instances.Source: §1 · p. 1
- Fine-tuning can overfit when each class has only 1 or 5 labeled examples, while meta-learning methods rely on an effective pretrained embedding network.Source: §1 · p. 1
- 69.69% vs. 68.14%: the proposed method achieves the reported state-of-the-art result in cross-domain few-shot learning.The introduction also reports nearly 3% and 4% gains on MiniImageNet 1-shot and 5-shot tasks, and nearly 9% and 3% on CUB, respectively.Source: §1 · p. 1
- The authors propose a much larger embedding network trained with self-supervised learning and report improvements over baselines on two common datasets.They also report gains from pretraining with more unlabeled data: 15% and 13% across two MiniImageNet tasks, and 15% and 8% across two CUB tasks.Source: §1 · p. 1
2. RELATED WORK
Prior work addresses few-shot learning through robustness improvements, meta-learning, and self-supervised representations; this paper combines SSL pretraining with episodic meta-learning.
2. RELATED WORK
- Other robustness efforts include orthogonal feature splitting for feature selection and dense estimation [10] and Self-Jig image augmentation using labeled or unlabeled synthesized images.Source: §2 · p. 1
- Meta-learning approaches seek robust adaptation with limited data, using gradient-based initialization, latent parameter embeddings, or learned feature spaces for nearest-neighbor classification.The cited examples include MAML, REPTILE [14], LEO, and embedding-based approaches.Source: §2 · p. 1
- The architecture first trains an embedding network by maximizing mutual information between augmented views, then compares episodic query embeddings with support-set centroids for prediction.Source: Figure 1 · p. 2
- Self-supervised pretraining learns representations without class labels, and the proposed method differs from prior combinations by using no supervised pretraining.Earlier work combined supervised and self-supervised losses, while another study found SSL embeddings comparable to supervised ones [19].Source: §2 · p. 2
3. METHOD
The method combines self-supervised pretraining of an embedding network with episodic meta-learning fine-tuning. Its pretraining learns representations from multiple views, and its meta-learning stage adapts them to few-shot classification tasks.
3. METHOD
- The method pairs self-supervised pretraining of a large embedding network with episodic meta-learning fine-tuning.The two-stage design uses self-supervised learning for the initial embedding and meta-learning in stage two.Source: §3 · p. 2, §3.2 · p. 3
3.1. Self-supervised learning stage
- AMDIM pretrains representations by maximizing mutual information across views of one image, using global and local features at 5 x 5 and 7 x 7 scales.The overall SSL loss sums three NCE losses: global-to-5 x 5, global-to-7 x 7, and 5 x 5-to-5 x 5 features.Source: §3.1 · p. 2
3.2. Meta-learning stage
- Meta-learning fine-tunes the stage-one embedding episodically on K-way C-shot tasks, using support and query sets.Each class descriptor is the mean embedding of its training samples; query predictions use Euclidean distances to class descriptors, converted to a distribution with softmax.Source: §3.2 · p. 3
4. EXPERIMENTAL RESULTS
The evaluation compares the method with few-shot baselines and tests transferability in a cross-domain setting. Results cover MiniImageNet, CUB, four additional datasets, and ablations of pretraining and meta-learning. [p23]
4.1. Datasets
- MiniImageNet has 60,000 images across 100 classes, split into 64 training, 16 validation, and 20 test classes; CUB is a fine-grained bird dataset, and four other datasets test cross-domain transfer.The additional datasets are CropDiseases, EuroSAT, ISIC, and ChestX; their similarity to MiniImageNet decreases across the set.Source: §4.1 · p. 3
4.2. Training Details
- The AMDIM setup uses AmdimNet with embedding dimension 1536, 128 × 128 unlabeled-image inputs for SSL, and 84×84 images during meta-learning.Mini80-SSL, CUB150-SSL, and Image1K-SSL are self-supervised variants; Mini80-SL and CUB150-SL are supervised variants.Source: §4.2 · p. 3
4.3.1. Standard few-shot learning evaluation
- On MiniImageNet, gains were 7.53% over ProtoNet+ [15] and 2.27% over LEO; on CUB, 1-shot accuracy was 71.85% and 5-shot accuracy was 84.29%.With Image900 SSL, MiniImageNet results averaged 16% improvement over baselines using the same amount of labelled data. On CUB, the reported gains were 20.54% over ProtoNet for 1-shot and 2.39% over DN4-Da for 5-shot.Source: Table 1 · p. 3, §4.3.1 · p. 4, Table 2 · p. 4
4.3.2. Cross-domain few-shot Learning
- Averaged results across four datasets and 5-way 5-shot, 20-shot, and 50-shot tasks show nearly 1.5% improvement over state-of-the-art results.The cross-domain tests use the Mini80-SSL embedding and the transductive setup from; the authors flag ISIC for further investigation.Source: §4.3.2 · p. 4, Table 3 · p. 4
4.4. Ablation Study
- The ablations report that SSL pretraining yields a more generalized embedding than supervised training of the large network, while meta-learning fine-tuning improves nearest-neighbour results.The supervised large network performs worse than methods using simple 4 Conv blocks; the authors attribute this to overfitting and difficulty adapting to unseen classes.Source: §4.4 · p. 4
- Adding more unlabeled data during SSL pretraining improves results; ImageNet-pretrained embeddings also transfer to CUB and four additional datasets.Reported gains are 15% and 13% across two test tasks with more unlabeled data, and 15% and 8% across both tasks when transferring ImageNet pretraining to CUB.Source: §4.4 · p. 4
5. CONCLUSION
The paper concludes that SSL can train a robust, more generalized and transferable embedding for few-shot image classification, followed by meta-learning fine-tuning. It reports that this approach outperforms baselines on two standard datasets and in a cross-domain setup, while leaving end-to-end training and other few-shot tasks as future directions.
5. CONCLUSION
- The proposed SSL-trained embedding, fine-tuned through meta-learning, outperforms baselines on two standard few-shot datasets and in a cross-domain setup.The authors identify combining the two stages end-to-end and testing few-shot detection as future directions.Source: §5 · p. 4