CCF-BSF: AF: Small: Algorithms for Interactive Learning

点击翻译按钮获取中文摘要

Machine learning classifiers are core components of many of the technologies we use routinely: search engines, speech recognition engines, language translators, assisted driving systems, and so on. These classifiers are typically built by a process of 'supervised learning', in which a computer is given a collection of (input, output) pairs that illustrate a desired behavior (e.g. if the input is this English sentence, the output should be this Spanish sentence) and is told to produce a function that replicates such behavior. This is a rigid form of learning that is known to suffer from a variety of fundamental hurdles; for instance, there are classes of concepts that cannot efficiently be learned in this way. This project will study how such hurdles can be overcome by moving to a more natural learning setup, in which the learning machine is allowed to interact with a human while learning, and receives feedback that is richer than just output values. This research has the potential to influence the way in which machine learning is performed and to broaden its scope of applicability. It is inherently multidisciplinary, and thus part of the project includes community-building activities that will bring together different groups of relevant researchers. There is also an educational component to the project, centered on bringing knowledge of algorithms and machine learning to various student groups that have traditionally been under-represented in STEM disciplines. Interactive learning is a field with great promise in which most of the work to date has consisted of one-off systems geared towards specific applications. This project will aim to bring rigor, formalism, and algorithms with provable guarantees to parts of this field that are currently lacking them.This project will aim to formalize forms of human feedback (to a learning machine) than are richer than those traditionally studied, such as: simple explanations (e.g. this bird is not a canary because it has the wrong type of beak); attention-focusing; and similarity judgments. The investigators will design algorithms that are able to use these kinds of feedback and have rigorous guarantees, both on correctness and on statistical rates of convergence. The project is particularly focused on overcoming fundamental hardness barriers in learning: learning concept classes that would be intractable to learn in the usual supervised framework; learning with dramatically fewer examples than would normally be needed; adapting to situations in which the distribution of the data is constantly shifting; and improving the results of unsupervised learning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.