This repo accompanies a workshop titled "Achieve More with Less using Active Learning" that I gave with Tal Perry on May 2018 as part of Data Science Summit Europe in Tel Aviv, Israel.
It contains a presentation and 2 notebooks, demonstrating the basics of Active Learning on the MNIST digits dataset and on a dataset of spam/ham phone text messages.
The introduction of large annotated datasets had a major contribution to the rise of deep neural networks and machine learning models in recent years. Still, in many fields obtaining a large set of labels is difficult, time consuming and expensive, whereas unlabeled data is cheap and abundant. Several works have shown that machine learning algorithms can achieve higher accuracy with fewer training labels using Active Learning - allowing them to choose the data from which they learn.
In this workshop, we'll provide a general introduction to Active Learning, including the theory and empirical evidence for its success and failures. We'll write code implementing leading Active Learning methods and use them on test sets for problems in Image Classification, Audio Classification and NLP.