Everything you need to know about Few-Shot Learning

The success of Deep Learning models in Computer Vision tasks like image classification, semantic segmentation, object detection, etc., is attributed to taking advantage of the vast amounts of labeled data used for training a network, a method called supervised learning. Although a large amount of unstructured data is available in this era of Information Technology, annotated data is challenging to come by.

Data Labeling takes the majority of the time devoted to a computer vision Machine Learning project, which is why it is also an expensive endeavor. Furthermore, in fields like healthcare, only expert doctors can categorize the data—for example, take a look at the following two images of cervical cytology: Can you definitively say which one is cancerous?

👁 Figure

Source: SIPaKMeD Dataset

Most untrained medical professionals will not know the answer: (a) is cancerous, while (b) is benign. So, data labeling is more difficult in such scenarios. At best, we would have only a handful of annotated samples, which is not nearly enough to train supervised learning models.

Also, newer data may become incrementally available over time—like data from a newly identified species of birds. Training a deep neural network on a large dataset consumes a lot of computational power (for example, ResNet-200 took about three weeks to train on 8 GPUs). Therefore, retraining the model to accommodate the newly available data is unfeasible in most scenarios.

This is where the relatively new concept of Few-Shot Learning comes in.