Friday, 24th November - 4pm Nuaha Dziri : Seminar

Title:  Faith and Fate: Limits of Transformers on Compositionality

Abstract: 

Transformer large language models have sparked admiration for their exceptional performance on tasks that demand intricate multi-step reasoning. Yet, these models simultaneously show failures on surprisingly trivial problems. This begs the question: Are these errors incidental, or do they signal more substantial limitations? In an attempt to demystify Transformers, in this talk, I will discuss the limits of these models across three representative compositional tasks—multi-digit multiplication, logic grid puzzles, and a classic dynamic programming problem. These tasks require breaking problems down into sub-steps and synthesizing these steps into a precise answer. We formulate compositional tasks as computation graphs to systematically quantify the level of complexity, and break down reasoning steps into intermediate sub-procedures. Our empirical findings suggest that Transformers solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching, without necessarily developing systematic problem-solving skills. To round off our empirical study, we provide theoretical arguments on abstract multi-step reasoning problems that highlight how Transformers’ performance will rapidly decay with increased task complexity.

Bio: 

Nouha Dziri is a research scientist at the Allen Institute for AI working with Yejin Choi and the Mosaic team. Prior to this, she earned her PhD from the University of Alberta and the Alberta Machine Intelligence Institute. Currently, her work revolves around three main axes: 1) Understanding the limits of Transformers and their inner workings. 2) Building smaller LMs that can learn more efficiently, and 3) better aligning LMs with human values and ethical principles. She has worked at Google Research, Microsoft Research, and Mila. Her work has been published in top-tier venues including Neurips, ICML,TACL, ACL, NAACL, and EMNLP. She actively serves as a reviewer for NLP conferences, journals, and workshops and was recognized among the best reviewers at ACL 2021. 

 vCal  iCal