5/27/2021 2:19:49 PM
Discover few-shot learning
Machine learning, a branch of artificial intelligence, has grown significantly in recent years. This rapid development is due to several factors, amongst them are improvements in algorithms and learning models both quantitatively and qualitatively, the increase of computing capability of machines, and the availability of big data. Yet collecting large amounts of data appropriately to increase the accuracy of predicting models is not always easily achieved. Few-shot learning algorithms are successful in making predictions using limited data only.[1]
Humans can recognize an object class by observing just a few instances of it; but when it comes to machines and machine learning techniques, they typically require thousands of examples to be given in the dataset. Few-shot learning (FSL), also known as low-shot learning (LSL), aims to solve this problem and achieve a human-like performance, by using only a few samples instead of large datasets. [2]
One-shot learning is also applicable, where one training image or sample is merely used. In addition to few-shot learning, there is zero-shot learning which aims to predict without being exposed to any prior example. Mutual applications of few-shot learning and zero-shot learning include image classification, image generation, object detection, semantic segmentation, and natural language processing (NLP). [3,4]
The main advantage of few-shot learning techniques is the elimination of the need for large amounts of data. When we have too many samples in our dataset, adding features for each task gets extremely difficult. By reducing the volume of our data, the cost of collecting and labeling them is significantly reduced. Another consequence is the reduction of the dimensionality of the dataset, which in turn results in lowering the computational costs. [4]
Therefore, few-shot learning can create more general models rather than highly specialized ones, which makes the models more robust. Also, when we have very few samples to train the model, rare illnesses for example, few-shot learning is expected to be very useful. In other words, a flexible model replaces the large volumes of data required, making the learning process more human-like. [3]
Applications of few-shot learning in medicine
While deep learning systems have provided breakthroughs in several medical domain tasks, they are still limited by the problem of dependency on the availability of training data. To counter this limitation, there is active research ongoing in few-shot learning. Few shot learning algorithms aim to overcome the data dependency by exploiting the information available from a tiny amount of data. In medical imaging, due to some diseases' rare occurrence, there is often a limitation on the available data. As a result, to which few-shot learning algorithms' success can prove to be a significant advancement. In this chapter, the background and working of few-shot learning algorithms are explained. The problem statement for few-shot classification and segmentation is described. There is then a detailed study of the problems faced in medical imaging related to limited data availability. After establishing context, the recent advances in applying few-shot learning to medical imaging tasks such as classification and segmentation are explored. The results of these applications are examined with a discussion on their future scope.
Rapid and accurate classification of medical images plays an important role in medical diagnosis. Nowadays, for medical image classification, some methods are based on machine learning, deep learning, and transfer learning. However, these methods may be time-consuming and not suitable for small datasets. We propose a novel method that combines a few-shot learning method and attention mechanism based on these limitations.
Regarding the use of this Few-shot Learning method in medical sciences, there are projects that we mention two of them.
Diagnosing the presence of cancer in the organs of the body from radiographic images is a complex task for two reasons: first, because the naked eye of a specialist doctor is not always very accurate and the possibility of errors in human diagnosis is high.
In diagnostics using machine learning methods and computer systems, large datasets were previously needed to perform modeling for diagnosis.
But in the article [5] Scientists were able to use the Few-shot learning method and by collecting about 20 Data/Class to achieve 90% accuracy in the diagnosis of breast and lung cancer tissue.
Or in another article [6], using the same Few-shot learning method, ophthalmologists were able to examine the presence of glaucoma in the eye by having a small database of eye images.
By: Ehsan Khormali - Mehdi Mohebali Zadeh
References:
1- https://medium.com/quick-code/understanding-few-shot-learning-in-machine-learning-bede251a0f67
2- https://www.borealisai.com/en/blog/tutorial-2-few-shot-learning-and-meta-learning-i/
3- https://www.unite.ai/few-shot-learning/
4- https://research.aimultiple.com/few-shot-learning/#differences-between-few-shot-learning-and-zero-shot-learning
5- Medela, Alfonso, et al. "Few shot learning in histopathological images: reducing the need of labeled data on biological datasets." 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). IEEE, 2019.
6- Kim, Mijung, Jasper Zuallaert, and Wesley De Neve. "Few-shot learning using a small-sized dataset of high-resolution fundus images for glaucoma diagnosis." Proceedings of the 2nd International Workshop on Multimedia for Personal Health and Health Care. 2017.