Reminder, this is happening today. Note, refreshments will not be provided.
On Sun, Dec 10, 2023 at 9:08 AM Sagie Benaim sagie.benaim@mail.huji.ac.il wrote:
Dear all,
Next week, we have the pleasure of having Dr. Amir Feder give a talk in the colloquium.
The seminar will be held on Monday, December 18th at 14:00. Location: C220.
The title, abstract and bio appear below.
Looking forward to seeing you, Sagie and Liat
*Title*: Causally-driven ML for Text
*Abstract:* A fundamental goal of scientific research is to learn about causal relationships, yet this aspect remains underrepresented in machine learning. Recent advances, particularly with large language models (LLMs), have primarily focused on leveraging neural architectures to extract correlations from extensive datasets. However, despite their success, correlational predictive models can be untrustworthy: they rely on shortcuts in the data, leading to errors when applied to out-of-distribution settings; and their representation of text lacks interpretability, rendering them unsuitable for scientific inquiry.
In this talk, I will show how a causal perspective can mitigate these shortcomings. I will show how by understanding causal relationships in data we can reduce reliance on shortcuts, and present causally-driven methods that leverage causal structures to perform better out-of-distribution. Then, I will briefly describe how to build probabilistic models of text that reveal interpretable latent variables tailored for estimating causal effects. Finally, I discuss my vision for building science-ready text models.
*Bio*: Amir Feder is a postdoctoral fellow at the Columbia University Data Science Institute, working with David Blei. He is currently also a visiting faculty researcher at Google Research. He works in the field of machine learning and causal inference, with a focus on text data. His research develops methods that integrate causality into natural language processing (NLP) to improve the reliability of NLP systems, and to facilitate scientific inquiry with text data. He was a co-organizer of the First Workshop on NLP and Causal Inference (CI+NLP) at EMNLP 2021, the Tutorial on Causality for NLP at EMNLP 2022, and the Workshop on Spurious Correlations, Invariance and Stability at ICML 2023.
Before joining Columbia, Amir received his PhD from the Technion, where he was advised by Roi Reichart and worked closely with Uri Shalit. In a previous (academic) life, he was an economics, statistics and history student at Tel Aviv University, the Hebrew University of Jerusalem and Northwestern University.