[Paper Reading] A CLOSER LOOK AT FEW-SHOT CLASSIFICATION
Abstract
Present
- A consistent comparative analysis of several representative few-shot
classification algorithms, with results showing that deeper backbones significantly reduce the performance differences among methods on datasets with limited domain differences - A modified baseline method that surprisingly achieves competitive performance when compared with the state-of-the-art on both the mini ImageNet and the CUB datasets
- A new experimental setting for evaluating the cross-domain generalization ability for few-shot classification algorithms
Introduction
- The strong performance of deep learning heavily relies on training a network with abundant labeled instances with diverse visual variations
- High cost of data
- Humans could learn with few data
- Few-shot classification: learning to generalize to unseen classes during training
- One promising solution: meta-learning paradigm where transferable knowledge is extracted and propagated from a collection of tasks to prevent overfitting and improve generalization
- Another solution: directly predicting the weights of the classifiers for novel classes
- Limitation: the discrepancy of the implementation details and lack of domain shift between the base and novel classes
Contributions
- Provide a unified testbed for several different few-shot classification algorithms for a fair comparison. Their empirical evaluation results reveal that the use of a shallow backbone commonly used in existing work leads to favorable results for methods that explicitly reduce intra-class
variation. Increasing the model capacity of the feature backbone reduces the performance gap between different methods when domain differences are limited - Show that a baseline method with a distance-based classifier surprisingly achieves competitive performance with the state-of-the-art meta-learning methods on both mini-ImageNet and CUB datasets
- Investigate a practical evaluation setting where base and novel classes are sampled from different domains. They show that current few-shot classification algorithms fail to address such domain shifts and are inferior even to the baseline method, highlighting the importance of learning
to adapt to domain differences in few-shot learning
Related Work
Initialization based methods
- Tackle the few-shot learning problem by “learning to fine-tune”
- Good model initialization
- Learning an optimizer
Distance metric learning based methods
- Address the few-shot classification problem by “learning to compare”
- If a model can determine the similarity of two images, it can classify an unseen input image with the labeled instances
Hallucination based methods
- Directly deal with data deficiency by “learning to augment”
- Learns a generator from data in the base classes and use the learned generator to hallucinate new novel class data for data augmentation
Domain adaptation
- Aim to reduce the domain shifts between source and target domain as well as novel tasks in a different domain
Overview of Few-shot Classification Algorithms
Baseline
Baseline++
Meta-learning Algorithms
Experimental Results
Scenarios/Datasets:
- Generic object recognition/mini-ImageNet
- Fine-grained image classification/CUB-200–2011
- Cross-domain adaptation/mini-ImageNet →CUB
Evaluation Using Standard Settings
Conclusion
They have investigated the limits of the standard evaluation setting for few-shot classification. Through comparing methods on a common ground, their results show that the Baseline++ model is competitive to state of art under standard conditions, and the Baseline model achieves competitive performance with recent state-of-the-art meta-learning algorithms on both CUB and mini-ImageNet benchmark datasets when using a deeper feature backbone. Surprisingly, the Baseline compares favorably against all the evaluated meta-learning algorithms under a realistic scenario where there exists domain shift between the base and novel classes.
Reference
Chen, Wei-Yu, et al. “A closer look at few-shot classification.” arXiv preprint arXiv:1904.04232 (2019).