Shourya Roy – from Labs to the Real World: Transfer Learning in Practice

798 views · Published 29 July 2016 · 49:29 · Indexed 27 October 2026

Channel: Hasgeek TV · 2016 · Science & Technology

Watch on YouTube

Traditional supervised learning models’ performances degrade if “nature” of test samples differ from that of training samples. For example, a classifier built to discriminate between “books” with positive, negative and neutral reviews when applied to discriminate between “kitchen products” into the same set categories, its performance drops. This relates to one of the fundamental probably approximately correct (PAC) assumptions that the training and test samples come from the same distribution. Consequently it leads to the practical implication that supervised models need to be provided with (enough number of) training samples from the domain where it is expected to be applied. This leads to laborious, tedious and ongoing labeling exercise limiting scalability and fast deployment of supervised algorithms.

In real life analytics applications, building models from scratch for every new domain hinders large-scale adoption of supervised statistical learning based analytics applications. Transfer learning techniques allow domains, tasks, and distributions used in training and testing to be different, thus reducing the requirement for labelled data. However, brute force techniques suffer from the problem of negative transfer, and we need to judge when and how much to transfer.

More from this channel