Active Learning Methods for Training Datasets

Active learning is a specific instance of supervised machine learning. The underlying principle of active learning is that, an ML algorithm might potentially attain a higher degree of accuracy with a reduced number of training labels, if it could choose which data it wanted to learn from. Using a modest amount of data with labels, the model is trained repeatedly for predicting labels of data that isn't labeled. Next, the data points selected for categorization are those about which the algorithm is least certain. This iterative method is carried out till the required degree of precision is attained. By actively selecting the important data points, this method reduces the size of the dataset used for training while creating a high-performance classifier. Moreover, it also reduces the cost and time of dataset labeling while also being able to attain a high degree of precision. This paper examines the application of active learning using the Fashion MNIST, EuroSAT, and CIFAR10 datasets. The output of different active learning algorithms, such as Margin Sampling, Entropy Sampling, and Least Confidence are compared. Furthermore, the model's performance over random sampling is also demonstrated. It is seen that the accuracy obtained by using Active learning is much higher as compared to random sampling.