We use deep learning algorithms mostly to classify images amongst a huge collection of images/videos. However, Few Shot Learning (FSL) is another set of algorithms which helps to build machine learning models using very less training samples. If configuration is done properly, one shot or few shot learning can be used for facial recognition and passport checks. FSL is different from standard supervised learning.
What is Few-Shot Learning?
Shot means examples available for training a data model. One shot learning thus implies one example available, few shot denotes few samples and so on. The main idea behind this technique is to predict similarity. After completion of training, the trained function can be used on unseen queries to make predictions. The similarity function compares the given query with all the samples in the support set and calculates a similarity score. Hence the sample with the highest score is the final prediction.
Approach for Few-Shot Learning
Few Shot Learning can fit in either of the three categories. Data-level Approach, Parameter-level Approach, Metric-based Approach.
Data-level Approach : Since availability of training data is very less for FSL, hence, data augmentation can help create more data for classification purpose.
Parameter-level Approach involves meta-learning. It teaches a model to differentiate which attributes are important for a ML task.
In Metric-based Approach, the algorithms uses distance metrics to make comparison between samples in data set. For image classification, it implies segregating images based on similarity of superficial characteristics.
Applications of Few Shot Learning
As Few shot Learning involves minimal datasets and low costs are enough, it can be applied in many segments such as healthcare, computer vision, robotics, natural language processing.
Healthcare: FSL can help in reducing the amount of data required to make effective predictions regarding drug discovery. This is an emerging field of AI healthcare.
Computer Vision: Image classification, character recognition are some of the many help that FSL can provide in Computer Vision.
Robotics: Robots can be trained using FSL technique. They can be trained to carry out actions, navigate the area around them, etc.
Natural Language Processing: Similarly, translation, sentence completion are some tasks which can be accomplished.
End Notes
Thus, we can say that few shot learning refers to a variety of algorithms which can build a model using a very small amount of training data. It is a boon for cases where the training can be compromised due to scarcity of data. Another advantage being its low cost, time and data management.
