Machine Learning works when we have a lot of labeled data. However, in many realistic settings we do not have enough training data. In most cases this is due to semantic shift (a shift in the labels space Y) or domain shift (where the domain X of the target is different from the domain for which we have training data) but can also be due to the complexity and compositionality of the task. Some examples for these setting are:
- Variable label space: imagine that you classify documents into a set of topical labels, and then want to use a different label space. Or that you classify entities into their semantic types, and then want to update the set of semantic types used.
- Domain adaptation: you train a model on news data but you want to use it on email data, where you don’t have training data
- Low Resource Languages: only 30 languages (out of around 3,500 written languages) have annotated data for basic tasks such as named entity recognition. How can we develop basic NLP tools for other languages?
- Complex tasks: many natural language understanding decisions are “one-in-a-million” – they are very sparse; how can we learn models for these?
And, of course, similar challenges exist in computer vision and other sub areas of AI.
The goal of this class is to define and understand the space of Learning in Low Labels Settings – understand the problems and the methods that have been studied for these setting. We will do this mostly in the context of natural language understanding with, possibly, some digressions to computer vision.
We will consider methods such as
- few/zero-shot setting
- semi-supervised and transductive learning
- self-supervised learning
- the use of incidental supervision signals
- transfer learning
- adaptation methods
And do it in the context of multiple tasks.
You will read, present and discuss papers, and work on two projects. A small, well-defined one, in the first third of the semester, and a large and open ended one in the rest of the semester.
|Feb 15, 2021||First Critical Survey Due|
|Mar 8, 2021||Second Critical Survey Due|
|Mar 15, 2021||Project 1 Paper Submission Deadline and Presentation|
|Mar 22, 2021||Project 2 Proposal Due|
|Mar 29, 2021||Third Critical Survey Due|
|Apr 5, 2021||Project 2 Progress Report and Brief Presentation|
|Apr 19, 2021||Fourth Critical Survey Due|
|Apr 26, 2021||Project 2 Final Presentation|
|May 5, 2021||Project 2 Due|
Machine Learning class; CIS 419/519/520 or equivalent. NLP: Knowledge of NLP (equivalent to a basic Computational Linguistics/NLP class).
Time and Location
Synchronously via Zoom