In few-shot classification, performing well on a testing dataset is a challenging task due to the restricted amount of labelled data available and the unknown distribution. Many previously proposed techniques rely on prototypical representations of the support set in order to classify a query set. Although this approach works well with a large, in-domain support set, accuracy suffers when transitioning to an out-of-domain setting, especially when using small support sets. To address out-of-domain performance degradation with small support sets, we propose Masked Embedding Modeling for Few-Shot Learning (MEM-FS), a novel, self-supervised, generative technique that reinforces few-shot-classification accuracy for a prototypical backbone model. MEM-FS leverages the data completion capabilities of a masked autoencoder to expand a given embedded support set. To further increase out-of-domain performance, we also introduce Rapid Domain Adjustment (RDA), a novel, self-supervised process for quickly conditioning MEM-FS to a new domain. We show that masked support embeddings generated by MEM-FS+RDA can significantly improve backbone performance on both out-of-domain and in-domain datasets. Our experiments demonstrate that applying the proposed technique to an inductive classifier achieves state-of-the-art performance on mini-imagenet, the CVPR L2ID Classification Challenge, and a newly proposed dataset, IKEA-FS. We provide code for this work at https://github.com/Brikwerk/MEM-FS.

Medical imaging tasks are very challenging due to the lack of publicly available labeled datasets. Hence, it is difficult to achieve high performance with existing deep learning models as they require a massive labeled dataset to be trained effectively. An alternative solution is to use pre-trained models and fine-tune them using a medical imaging dataset. However, all existing models are pre-trained using natural images, which represent a different domain from that of medical imaging; this leads to poor performance due to domain shift. To overcome these problems, we propose a pre-trained backbone using a collected medical imaging dataset with a self-supervised learning tool called a masked autoencoder. This backbone can be used as a pre-trained model for any medical imaging task, as it is trained to learn a visual representation of different types of medical images. To evaluate the performance of the proposed backbone, we use four different medical imaging tasks. The results are compared with existing pre-trained models. These experiments show the superiority of our proposed backbone in medical imaging tasks.

Image classification in real-world applications is a challenging task due to the lack of labeled data. Many few-shot learning techniques have been developed to tackle this problem. However, existing few-shot learning techniques will fail if new classes are added to the model without any labeled data (out-of-distribution). The existing techniques will classify the new classes as one of the existing classes, and will not be able to detect them as new classes. Also, if all data is unlabeled, existing few-shot learning techniques will not work (because existing techniques are supervised learning). This paper proposes a novel few-shot learning network (KTNet) that can learn from unlabeled data and assign a pseudo label to these data. These pseudo labeled data will be either added to the existing labeled data (in-distribution) to increase the number of shots or added as new classes if the data is out-of-distribution. The proposed KTNet technique can work in all cases, 1) a small amount of labeled data exist for all classes, 2) labeled data exist for a subset of all classes, and 3) if existing data for all classes is unlabeled. KTNet is evaluated using two benchmark datasets (mini-imagenet and fewshot-CIFAR). The results show that the proposed network outperforms state-of-the-art models in both datasets with respect to classification accuracy. Also, KTNet is better than existing techniques at detecting and clustering the out-of-distribution classes.

Few-shot learning aims to identify unseen classes with limited labelled data. Recent few-shot learning techniques have shown success in generalizing to unseen classes; however, the performance of these techniques has also been shown to degrade when tested on an out-of-domain setting. Previous work, additionally, has also demonstrated increasing reliance on supervised finetuning in an off-line or online capacity. This paper proposes a novel, fully self-supervised few-shot learning technique (FSS) that utilizes a vision transformer and masked autoencoder. The proposed technique can generalize to out-of-domain classes by finetuning the model in a fully self-supervised method for each episode. We evaluate the proposed technique using three datasets (all out-of-domain). As such, our results show that FSS has an accuracy gain of $1.05\%$, $0.12\%$, and $1.28\%$ on the ISIC, EuroSat, and BCCD datasets, respectively, without the use of supervised training.

Medical imaging tasks are very challenging due to the lack of publicly available labeled datasets. Hence, it is difficult to achieve high performance with existing deep-learning models as they require a massive labeled dataset to be trained effectively. An alternative solution is to use pre-trained models and fine-tune them using the medical imaging dataset. However, all existing models are pre-trained using natural images, which is a completely different domain from that of medical imaging, which leads to poor performance due to domain shift. To overcome these problems, We propose a backbone pre-trained using the collected medical imaging dataset with a self-supervised learning technique called Masked autoencoder. This backbone can be used as a pre-trained model for any medical imaging task, as it is trained to learn a visual representation of different types of medical images. To evaluate the performance of the proposed backbone, we used four different medical imaging tasks. The results are compared with existing pre-trained models. These experiments show the superiority of our proposed backbone in medical imaging tasks. The source code is available on github.com/islamosmanubc/ MedMAE.