Maxime Peyrard
@peyrardmax.bsky.social
Junior Professor CNRS (previously EPFL, TU Darmstadt) -- AI Interpretability, causal machine learning, and NLP. Currently visiting @NYU
https://peyrardm.github.io
https://peyrardm.github.io
We find a lot of identifiability issues:
- Multiple explanatory algorithms exists
- Even for one algorithm, there are many localizations in the network
Identifiability problems remain across scenarios: changing levels of over-parametrization, progress in training, multi-tasks, model size.
- Multiple explanatory algorithms exists
- Even for one algorithm, there are many localizations in the network
Identifiability problems remain across scenarios: changing levels of over-parametrization, progress in training, multi-tasks, model size.
April 21, 2025 at 1:52 PM
We find a lot of identifiability issues:
- Multiple explanatory algorithms exists
- Even for one algorithm, there are many localizations in the network
Identifiability problems remain across scenarios: changing levels of over-parametrization, progress in training, multi-tasks, model size.
- Multiple explanatory algorithms exists
- Even for one algorithm, there are many localizations in the network
Identifiability problems remain across scenarios: changing levels of over-parametrization, progress in training, multi-tasks, model size.
Mechanistic Interpretability aims to produce statements like: "Model M solves task T by doing X."
To do so, many causal manipulations are performed to validate an explanation. But what if (many) other, incompatible explanations also pass the causal tests?
To do so, many causal manipulations are performed to validate an explanation. But what if (many) other, incompatible explanations also pass the causal tests?
April 21, 2025 at 1:52 PM
Mechanistic Interpretability aims to produce statements like: "Model M solves task T by doing X."
To do so, many causal manipulations are performed to validate an explanation. But what if (many) other, incompatible explanations also pass the causal tests?
To do so, many causal manipulations are performed to validate an explanation. But what if (many) other, incompatible explanations also pass the causal tests?
Our paper "Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable?" will be presented at #ICLR2025!
It's also the first paper of my first PhD student, congrats @maximemeloux.bsky.social ! 🎉
blog: melouxm.github.io/MI-identifia...
An explanatory thread 🧵:
It's also the first paper of my first PhD student, congrats @maximemeloux.bsky.social ! 🎉
blog: melouxm.github.io/MI-identifia...
An explanatory thread 🧵:
April 21, 2025 at 1:52 PM
Our paper "Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable?" will be presented at #ICLR2025!
It's also the first paper of my first PhD student, congrats @maximemeloux.bsky.social ! 🎉
blog: melouxm.github.io/MI-identifia...
An explanatory thread 🧵:
It's also the first paper of my first PhD student, congrats @maximemeloux.bsky.social ! 🎉
blog: melouxm.github.io/MI-identifia...
An explanatory thread 🧵: