NLIP Seminar Series

Abstract: In this talk, I'll cover two recent papers for preference alignment: Odds-Ratio Preference Optimisation (ORPO, EMNLP 2024), discussing the role of the reference model for preference alignment (e.g. DPO, RLHF), and Margin-aware Preference Optimization (under review @ CVPR), thinking about the risks of reference...

Note this seminar has been rescheduled from its original date and will be taking place at 4 pm. Control theory is fundamental in the design and understanding of many natural and engineered systems, from cars and robots to power networks and bacterial metabolism. It studies dynamical systems—systems whose properties evolve...

We build neural networks in a modular and programmatic way using software libraries like PyTorch and JAX. But optimization theory has not caught up to the flexibility of this paradigm, and practical advances in neural net optimization are largely heuristics driven. In this talk we argue that, if we are to treat deep...

Abstract: Yiwen Chen's study looks at generating lyrics in Mandarin that match well with both the melody and the tonal contour of the language. The approach uses mBART and treats lyrics generation as a sequence-to-sequence (seq2seq) task. Instead of generating lyrics directly from the melody, which is the usual way, the...

Abstract: Current large language models (LLMs) predominantly use subword tokenization. They see text as chunks (called "tokens") made up of individual words, or parts of words. This has a number of consequences. For example, LLMs often struggle with seemingly simple tasks involving character-level knowledge, like counting...

Recent chatbots have amazed everyone with their human-like language output. However, their relationship to research in the linguistics is opaque; even their inventors do not fully understand why they are so successful. Further, when probed in depth, some of their outputs are less human-like than first impressions would...

In this talk, Max Bartolo will share a brief overview of the critical role human feedback plays in enhancing Large Language Model (LLM) performance and aligning model behaviours to human expectations. We will delve into key aspects of human feedback, examining some of its requirements, benefits, and challenges. We will...

Abstract: State-of-the-art foundation models (FMs) process information as a sequence of internal representations; however, the length of this sequence is fixed and entirely determined by tokenization. This essentially decouples representation granularity from information content, which exacerbates the deployment costs of...

The statistical properties of language and how they may be used in language processing and language acquisition have been studied for many decades. Recently, large language models have demonstrated striking language-learning capabilities, providing evidence for the “richness” of the linguistic stimulus, but are often...

Truth-conditional semantics has been successful in explaining how the meaning of a sentence can be decomposed into the meanings of its parts, and how this allows people to understand new sentences. In this talk, I will show how a truth-conditional model can be learnt in practice on large-scale datasets of various kinds (...