Abstract: Leila Zilles
Mentor: Eric Eaton
In machine learning, classification tasks are usually learned independently and from scratch, drawing upon only the knowledge specified in the task domain. This approach is successful for large, labeled datasets, but can require a great deal of time and computation; furthermore, datasets with very few labels may result in highly inaccurate classifiers, depending on how the data is distributed. However, if similar tasks has been performed in the past, it may be possible to use information from these tasks to learn more rapidly and reduce generalization error overall. Transfer learning attempts to incorporate past knowledge into the learning process in order to enhance it in some way; the data provided in a typical transfer scenario consists of one or more source tasks with past knowledge, and a target task to be learned. Additionally, there may be some cases in which more labels are available for a dataset, but obtaining them comes at a cost. Active learning allows the system to query the user for labels when needed and addresses the problem of determining what query would provide the most information while incurring the least amount of cost.
Since transfer learning is often applied in situations where there is a lack of labeled data to learn from, it would be useful to somehow incorporate active learning into transfer learning to improve the transfer process. An open issue in transfer learning is detecting and repairing transfer failure, which occurs when the transfer information is not relevant and transfer has a negative effect on the classifier, resulting in worse performance than if the classifier had been trained only on the target task only. We will be using active learning techniques to aid in selecting good source tasks to transfer in order to minimize the probability of negative transfer, as well as validating the transfer process to ensure that transfer is having a positive effect on the target classifier.