| Description |
Machine learning is the science of building predictive models from data that automatically improve based on past experience. To learn these models, traditional learning algorithms require labeled data. They also require that the entire dataset fits in the memory of a single machine. Labeled data are available or can be acquired for small and moderately sized datasets but curating large datasets can be prohibitively expensive. Similarly, massive datasets are usually too huge to fit into the memory of a single machine. An alternative is to distribute the dataset over multiple machines. Distributed learning, however, poses new challenges as most existing machine learning techniques are inherently sequential. Additionally, these distributed approaches have to be designed keeping in mind various resource limitations of real-world settings, prime among them being intermachine communication. With the advent of big datasets machine learning algorithms are facing new challenges. Their design is no longer limited to minimizing some loss function but, additionally, needs to consider other resources that are critical when learning at scale. In this thesis, we explore different models and measures for learning with limited resources that have a budget. What budgetary constraints are posed by modern datasets? Can we reuse or combine existing machine learning paradigms to address these challenges at scale? How does the cost metrics change when we shift to distributed models for learning? These are some of the questions that have been investigated in this thesis. The answers to these questions hold the key to addressing some of the challenges faced when learning on massive datasets. In the first part of this thesis, we present three different budgeted scenarios that deal with scarcity of labeled data and limited computational resources. The goal is to leverage transfer information from related domains to learn under budgetary constraints. Our proposed techniques comprise semisupervised transfer, online transfer and active transfer. In the second part of this thesis, we study distributed learning with limited communication. We present initial sampling based results, as well as, propose communication protocols for learning distributed linear classifiers. |
