The Calcutta Mathematical Society published two important pioneer papers in information geometry in the same issue, vol. 37, in 1945: - The Fisher-Rao geometry celebrated paper of CR Rao - The dualistic structure of connections of RN Sen PDF of Vol 37 : ia802206.us.archive.org/7/items/dli.ca…

Excited to share a paper that's been in the works for a while, with Davin Choo, Arnab Bhattacharyya, David Sontag. "Probably approximately correct high-dimensional causal effect estimation given a valid adjustment set" arxiv.org/abs/2411.08141

Minimizing Kullback-Leibler divergence can be interpreted as an information projection wrt to Fisher orthogonality and exponential or mixture connection. Uniqueness of projection proof may be proved with a generalization of the Pythagoras' theorem!

youtube.com/watch?v=IFgHw_…

An advent calendar of some of my favourite TCS/Maths talks. Day #1: Avi Wigderson on Reading Alan Turing. It is a gem of a talk, full of insights about Turing's work, writing style, and influences on mathematics and computer science. Pure joy! youtube.com/watch?v=_Uk_ic…

You have a complicated system with several high-dimensional variables, such as image and text data. How can you systematically answer ANY causal question from observational data? #NeurIPS2024

Book #OTD "Statistical Structure of Quantum Theory" by Alexander S. Holevo (2016 Claude E. Shannon Awardee) 1/2

Compositional Risk Minimization 📜 arxiv.org/abs/2410.06303 💬 Compositional Learning Workshop (15th Dec) TLDR: Provable method for extrapolating classifiers to novel combinations of attributes

Excited to present Axiomatic Training at #NeurIPS2024, a new paradigm to teach causal reasoning to language models! I try to summarize what LLM systems can do today and what new training paradigms we need to improve their causal reasoning. Slides: amitsharma.in/talk/teaching-…

Catch our oral presentation, "Causal Order: Key to Leverage Imperfect Experts," at C♥️LM Workshop 2024, NeurIPS by Amit Sharma! We present a novel strategy leveraging imperfect experts (humans, LLMs) for accurate causal order discovery, overcoming pairwise approach limitations.

The first paper studies functions: linear, sparse linear, decision trees and relu. These are continuous space functions. But the space of functions that can be described in words is much more complex. For example, "translate", or "summarise".. these are much more complex.

Excited to share our work, "Causal Order: Key to Leveraging Imperfect Experts in Causal Inference," is accepted at #ICLR2025! 🎉 We propose a triplet-based querying strategy to leverage imperfect expert (LLM/humans) knowledge for causal discovery. openreview.net/forum?id=9juye…)

📢 Excited to announce the #ICML2025 workshop on *Scaling Up Intervention Models (SIM)*! Let’s bring together state-of-the-art ideas on modeling novel interventions and distribution shifts. :) 🙌🏻 Submissions are welcome! Link: sites.google.com/view/sim-icml2…

Excited to present our work in ICLR this week on LLMs and Causality! We try to answer the question: How can imperfect experts like Humans or LLMs be optimally used for causal discovery? Amit Sharma and I will be @ Hall 3 + Hall 2B #459 (24th April, 3PM - 5:30 PM)! Drop by!!

I will be presenting 👇work at #NAACL2025 tomorrow (May 2) from 12 pm in Ballroom A. Please stop by if curious about inductive biases in transformers, generalization, and applying Bayesian models of cognition for understanding language models.

Episode 2 of Causal Foundations will be released on Monday. Subscribe to my channel be notified when it is posted. youtube.com/@dagophile?si=…

Mathematician Eva Miranda expounds on undecidability in fluid mechanics. This is fun: It's about rubber duckies drifting in the ocean. youtube.com/watch?v=cgNpC-…

Happy to share that FlawedFictions is now accepted in CoLM 2025 (Conference on Language Modeling)! Looking forward to presenting this in Montreal later this year. Huge thanks to my wonderful collaborators Melanie Sclar and tsvetshop.

3/5 "Causal Abstraction Inference under Lossy Representations" (w/ Kevin Xia) Thu, 11 AM (East Ballroom, 2108) Link: causalai.net/r124.pdf Representation learning is a key component of modern ML, especially in high-dimensional settings. A central insight from

One question I’ve received a few times, and would like to clarify about this work (causalai.net/r115.pdf), is: why do we need identification and the ctf-calculus? Isn’t the do-calculus enough? The answer to the first question is that identification is essential: estimating a