Bridging the Gap Between Few-Shot and Many-Shot Learning via Distribution Calibration

Yang, Shuo; Wu, Songhua; Liu, Tongliang; Xu, Min

doi:10.1109/tpami.2021.3132021

“…To evaluate the performance of the proposed information fusion rectification, the results of 5-way 1-shot and 5way 5-shot classification are compared with related image augmentation and other state-of-the-art meta learning methods, including DC 13 , MAP 14 , CQD 15 , LEO 22 , SIB 23 , E 3 BM 24 , S2M2 25 ,and feature extractor of each method is set as WideResNet.…”

Section: Comparison Resultsmentioning

confidence: 99%

“…Zhang et al 12 utilize object detection algorithm to separate foreground and background of image, and then randomly combine the foreground and background of different images to generate more images. Distribution Calibration (DC) 13 uses distribution statistics (mean and covariance) of the base class data to calibrate the statistics of new class and extend the features of new class through sampling. Wu et al 14 propose to obtain the distribution statistics of new class through the maximum a posteriori (MAP) of base class and extend the features of new class through sampling.…”

Section: Data Augmentationmentioning

confidence: 99%

Feature augmentation based on information fusion rectification for few-shot image classification

Wang

¹

,

Tian

²

,

Fu

³

et al. 2022

Preprint

View full text Add to dashboard Cite

In the issue of few-shot image classification, due to lack of sufficient data, directly training the model will lead to overfitting. In order to alleviate this problem, more and more methods focus on non-parametric data augmentation, which uses the information of known data to construct non-parametric normal distribution to expand samples in the support set. However, there are some differences between base class data and new ones, and the distribution of different samples belonging to same class is also different. The sample features generated by the current methods may have some deviations. A new few-shot image classification algorithm is proposed on the basis of information fusion rectification (IFR), which adequately uses the relationship between the data (including the relationship between base class data and new ones, and the relationship between support set and query set in the new class data), to rectify the distribution of support set in the new class data. In the proposed algorithm, feature of support set is expanded through sampling from the rectified normal distribution, so as to augment the data. Compared with other image augmentation algorithms, the experimental results on three few-shot datasets show that the accuracy of the proposed IFR algorithm is improved by 0.93%-3.22%.

show abstract

“…The IFR based feature augmentation method contains three important module: the Tukey’s ladder of powers transformation module, the nearest neighbor prototypes matching module and the information fusion rectification module. The Tukey’s ladder of powers transformation has proved to be effective in reducing the deviation of the distributions by DC 24 , MAP 25 , CQD 26 . Hence, the Tukey’s ladder of powers transformation is adopted in IFR based feature augmentation method.…”

Section: Methodsmentioning

confidence: 99%

“…Zhang et al 23 utilize object detection algorithm to separate foreground and background of image, and then randomly combine the foreground and background of different images to generate more images. Distribution Calibration (DC) 24 uses distribution statistics (mean and covariance) of the base class data to calibrate the statistics of new class and extend the features of new class through sampling. Wu et al 25 propose to obtain the distribution statistics of new class through the maximum a posteriori (MAP) of base class and extend the features of new class through sampling.…”

Section: Introductionmentioning

confidence: 99%

Feature augmentation based on information fusion rectification for few-shot image classification

Wang

¹

,

Tian

²

,

Fu

³

et al. 2023

Sci Rep

1

0

View full text Add to dashboard Cite

In the issue of few-shot image classification, due to lack of sufficient data, directly training the model will lead to overfitting. In order to alleviate this problem, more and more methods focus on non-parametric data augmentation, which uses the information of known data to construct non-parametric normal distribution to expand samples in the support set. However, there are some differences between base class data and new ones, and the distribution of different samples belonging to same class is also different. The sample features generated by the current methods may have some deviations. A new few-shot image classification algorithm is proposed on the basis of information fusion rectification (IFR), which adequately uses the relationship between the data (including the relationship between base class data and new ones, and the relationship between support set and query set in the new class data), to rectify the distribution of support set in the new class data. In the proposed algorithm, feature of support set is expanded through sampling from the rectified normal distribution, so as to augment the data. Compared with other image augmentation algorithms, the experimental results on three few-shot datasets show that the accuracy of the proposed IFR algorithm is improved by 1.84–4.66% on 5-way 1-shot task and 0.99–1.43% on 5-way 5-shot task.

show abstract

“…Specifically, FSCC freezes the parameters of the feature extractor of the pre-trained model. Then, based on distribution calibration ( Yang et al, 2021b ), FSCC retrains the classifier with a limited number of labeled macromolecules from novel classes. Distribution calibration is a kind of domain adaption method ( Sun and Saenko, 2016 ).…”

Section: Introductionmentioning

confidence: 99%

FSCC: Few-Shot Learning for Macromolecule Classification Based on Contrastive Learning and Distribution Calibration in Cryo-Electron Tomography

Gao

¹

,

Zeng

²

,

Xu

³

et al. 2022

Front. Mol. Biosci.

2

0

View full text Add to dashboard Cite

Cryo-electron tomography (Cryo-ET) is an emerging technology for three-dimensional (3D) visualization of macromolecular structures in the near-native state. To recover structures of macromolecules, millions of diverse macromolecules captured in tomograms should be accurately classified into structurally homogeneous subsets. Although existing supervised deep learning–based methods have improved classification accuracy, such trained models have limited ability to classify novel macromolecules that are unseen in the training stage. To adapt the trained model to the macromolecule classification of a novel class, massive labeled macromolecules of the novel class are needed. However, data labeling is very time-consuming and labor-intensive. In this work, we propose a novel few-shot learning method for the classification of novel macromolecules (named FSCC). A two-stage training strategy is designed in FSCC to enhance the generalization ability of the model to novel macromolecules. First, FSCC uses contrastive learning to pre-train the model on a sufficient number of labeled macromolecules. Second, FSCC uses distribution calibration to re-train the classifier, enabling the model to classify macromolecules of novel classes (unseen class in the pre-training). Distribution calibration transfers learned knowledge in the pre-training stage to novel macromolecules with limited labeled macromolecules of novel class. Experiments were performed on both synthetic and real datasets. On the synthetic datasets, compared with the state-of-the-art (SOTA) method based on supervised deep learning, FSCC achieves competitive performance. To achieve such performance, FSCC only needs five labeled macromolecules per novel class. However, the SOTA method needs 1100 ∼ 1500 labeled macromolecules per novel class. On the real datasets, FSCC improves the accuracy by 5% ∼ 16% when compared to the baseline model. These demonstrate good generalization ability of contrastive learning and calibration distribution to classify novel macromolecules with very few labeled macromolecules.

show abstract

Bridging the Gap Between Few-Shot and Many-Shot Learning via Distribution Calibration

Cited by 19 publications

References 26 publications

Feature augmentation based on information fusion rectification for few-shot image classification

Feature augmentation based on information fusion rectification for few-shot image classification

Feature augmentation based on information fusion rectification for few-shot image classification

FSCC: Few-Shot Learning for Macromolecule Classification Based on Contrastive Learning and Distribution Calibration in Cryo-Electron Tomography

Contact Info

Product

Resources

About