While the unparalleled success of sophisticated learning algorithms (such as deep neural networks) has depicted commendable performance in several computer vision applications, training a robust machine learning model necessitates a large amount of hand-labeled training data, which is time-consuming and labor-intensive to acquire. This has motivated research in the field of weakly supervised learning, where the objective is to induce a robust machine learning model under the constraint that human annotation effort is expensive. Active Learning (AL) and Domain Adaptation (DA) are two popular learning paradigms which attempt to address the challenge of weakly supervised learning. AL algorithms automatically select the salient and exemplar instances from large amounts of unlabeled data; this not only tremendously reduces the human annotation effort in training an effective model, but also exposes the model to the informative samples in the underlying data population. DA algorithms address the problem of learning with weak supervision by utilizing abundant labeled data in one domain (source) to develop a model for a related domain of interest (target), where there is a paucity of labeled data, under the constraint of a probability distribution difference between the two domains. Of late, there has been a growing research interest in the problem of active domain adaptation (ADA), which attempts to combine the two learning paradigms. In ADA, we are given a large amount of labeled samples in the source domain, together with a small amount of labeled data and a large amount of unlabeled data in the target domain. The goal is three-fold: (i) addressing the disparity between the source and target domains; (ii) identifying the informative unlabeled samples in the target domain for manual annotation; and (iii) learning informative feature representations from the data using a deep neural network. AL, DA and ADA have been used with remarkable success in a variety of computer vision applications, such as object recognition and detection, image segmentation, image retrieval and medical imaging among others.

This tutorial will seek to present a comprehensive overview of active learning, domain adaptation and in particular, active domain adaptation, with a focus on computer vision applications, including historical perspectives, theoretical analysis and novel paradigms. The novelty of this tutorial lies in its focus on the emerging trends, algorithms and applications of these learning paradigms. It will aim at introducing concepts and open perspectives that motivate further work in this domain, ranging from fundamentals to applications and systems. We present this tutorial in 3 parts.

Part 1: Active Learning

We will introduce the concept of active learning (AL) and its origin in the field of education. The need for active learning will be motivated by highlighting the generation of enormous amounts of digital data (in the form of images, videos, text etc.) in the modern era and the scarcity of available human effort to hand-annotate the data for inducing machine learning models. After an overview of the basic concepts (we will delve into further details in Part 3, where we will discuss Active Domain Adaptation), we plan to cover a few recent research topics in Active Learning viz., AL with Imperfect Oracles, AL with Novel Query Types, AL for Open-set Classification.

Part 2: Domain Adaptation

We will introduce domain adaptation (DA) in the context of knowledge transfer paradigms like, multi-task learning, few-shot/zero-shot learning, incremental/continual learning, meta-learning and transfer learning. We will outline how domain adaptation is formulated and evaluated in the research community, and highlight popular datasets used for DA. In this module we will briefly discuss topics like, Feature and Pixel Alignment, DA in Unconstrained Label Spaces and Variats of DA. We will go into relevant details in the module on Active Domain Adaptation.

Part 3: Active Domain Adaptation

Research in active domain adaptation (ADA) is still in its nascent stage; however, it has been attracting significant research attention in recent years. We plan to dedicate about half of our time in this tutorial to this topic, due to its rising popularity in the computer vision community. We will discuss topics including, Conventional ADA, Active Source-free DA, Active Universal DA.