98 - Analyzing Information Flow In Transformers, With Elena Voita

98 - Analyzing Information Flow In Transformers, With Elena Voita

Author: Allen Institute for Artificial Intelligence December 9, 2019 Duration: 37:05
What function do the different attention heads serve in multi-headed attention models? In this episode, Lena describes how to use attribution methods to assess the importance and contribution of different heads in several tasks, and describes a gating mechanism to prune the number of effective heads used when combined with an auxiliary loss. Then, we discuss Lena’s work on studying the evolution of representations of individual tokens in transformers model. Lena’s homepage: https://lena-voita.github.io/ Blog posts: https://lena-voita.github.io/posts/acl19_heads.html https://lena-voita.github.io/posts/emnlp19_evolution.html Papers: https://arxiv.org/abs/1905.09418 https://arxiv.org/abs/1909.01380

While NLP Highlights is currently on hiatus, its archive remains a compelling snapshot of conversations from the front lines of computational linguistics. Produced by the Allen Institute for Artificial Intelligence, this science podcast carved out a space for deep, researcher-led discussions about natural language processing. Each episode functions as an informal seminar, where the people actively designing algorithms and pushing the field forward explain their work in their own words. You'll hear about the nuanced challenges behind making machines understand, generate, and reason with human language, from foundational theories to unexpected applications. The dialogue in this podcast often delves into the "why" behind the research, not just the results, offering clarity on complex topics like machine translation, sentiment analysis, or large language models. It’s a chance to listen as experts articulate their thought processes, debates, and moments of insight, with all the candidness that comes from a conversation between peers. The views shared are personal perspectives from the hosts and their guests, independent of their affiliated institutions. For anyone curious about how machines learn to parse meaning, the archived episodes of NLP Highlights provide a thoughtful and accessible entry point.
Author: Language: English Episodes: 100

NLP Highlights
Podcast Episodes
63 - Neural Lattice Language Models, with Jacob Buckman [not-audio_url] [/not-audio_url]

Duration: 30:03
TACL 2018 paper by Jacob Buckman and Graham Neubig. Jacob tells us about marginalizing over latent structure in a sentence by doing a clever parameterization of a lattice with a model kind of like a tree LSTM. This lets…
56 - Deep contextualized word representations, with Matthew Peters [not-audio_url] [/not-audio_url]

Duration: 30:06
NAACL 2018 paper, by Matt Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Chris Clark, Kenton Lee, and Luke Zettlemoyer. In this episode, AI2's own Matt Peters comes on the show to talk about his recent work on ELMo emb…