Supervisors

Suitable for

Abstract

We have a range of projects at the intersection of active learning/subset selection and large language models (LLMs). Active learning is a machine learning approach where the algorithm selectively queries a user or another information source to label new data points with the aim of reducing the amount of data that needs to be labelled for training a model. The motivation behind this approach lies in its efficiency and cost-effectiveness, especially in scenarios where labelling data or training is expensive or time-consuming. It is particularly relevant in the context of LLMs, where finetuning is computationally expensive. The intersection of active learning with LLMs opens new avenues for more efficient model training and fine-tuning, enabling these models to learn more effectively from smaller, well-chosen subsets of data. This area includes exploring how best to select these subsets, determining which examples are most beneficial for tasks like few-shot learning, and actively selecting demonstrations or test examples to optimize the model's performance and evaluation.

This research aims to improve the efficiency and effectiveness of large language models through advanced active learning and subset selection techniques. The key research question investigates how these methodologies can be best applied to reduce training resources while enhancing model performance in diverse NLP tasks. The projects target submission at the major ML conferences.

Goals:

To be determined based on the project chosen. These projects have been scoped and we will develop a full project plan with the student(s). An expected goal would to be produce research that is of publishable quality at a major ML conference.

References:

• Active Learning in LLMs (including active model selection and example selection)
  • https://arxiv.org/abs/2305.14264
  • https://arxiv.org/abs/2211.04486
  • https://arxiv.org/abs/2302.12246
• Active Learning:
  • Active Learning in DL:
    • https://arxiv.org/abs/1112.5745
    • https://arxiv.org/abs/1703.02910
    • https://www.cs.ox.ac.uk/people/yarin.gal/website/udl23/
  • Active Model Selection: https://arxiv.org/abs/1207.4138
  • Active Testing: https://arxiv.org/abs/2103.05331

 

 Pre-requisites: Machine Learning and Deep Learning