iiinova — where your digital self keeps evolving

Daniel Csillag

@dccsillag.xyz

888 followers 378 following 62 posts

Applied mathematician working on machine learning, statistics and compilers. Currently doing research at FGV EMAp. dccsillag.xyz

Posts Replies Media

Daniel Csillag @dccsillag.xyz · 04/02/2025

In this paper we propose strategic conformal prediction, a simple extension of the standard conformal prediction methods which ensures that the uncertainty quantification is robust to strategic alterations.

Main theorems of our work, establishing marginal and training-conditional guarantees under strategic alterations for our method.

One of our plots, showcasing the effectiveness of our method.

110

Daniel Csillag @dccsillag.xyz · 04/02/2025

Happy to announce that our paper, 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗰 𝗖𝗼𝗻𝗳𝗼𝗿𝗺𝗮𝗹 𝗣𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻, is now accepted to AISTATS 2025! Is your uncertainty quantification robust to people trying to break it? Ours is :)

Front page of our paper, 'Strategic Conformal Prediction'.

Abstract: When a machine learning model is deployed, its predictions can alter its environment, as better informed agents strategize to suit their own interests. With such alterations in mind, existing approaches to uncertainty quantification break. In this work we propose a new framework, Strategic Conformal Prediction, which is capable of robust uncertainty quantification in such a setting. Strategic Conformal Prediction is backed by a series of theoretical guarantees spanning marginal coverage, training-conditional coverage, tightness and robustness to misspecification that hold in a distribution-free manner. Experimental analysis further validates our method, showing its remarkable effectiveness in face of arbitrary strategic alterations, whereas other methods break.

110

Daniel Csillag @dccsillag.xyz · 21/11/2024

Anthropic published a paper telling people to use confidence intervals for evals. I now await for their next paper, which will explain multiple comparisons to the LLM people

110

Daniel Csillag @dccsillag.xyz · 04/09/2024

Gradient descent with random initialization avoids troublesome stationary points almost surely. #🧮

Let $f : RR^n -> RR$ be twice continuously differentiable and assume its gradient is Lipschitz.
Consider gradient descent with random initialization:

$ x_0 ~ nn(bold(0),I), wide wide x_(t+1) = x_t - eta nabla f(x_t). $

If the procedure converges to some point $macron(x)$ and $eta$ is sufficiently small, then, almost surely:

(i) It is a critical/stationary point: $nabla f(macron(x)) = 0$.
(ii) It is a second-order stationary point: the Hessian $nabla^2 f(macron(x))$ is positive-semidefinite.

This means that it will not converge to maxima, nor most other spurious stationary points.

So, even though GD only uses first-order information (the gradient), we get a second-order guarantee "for free!"

170

Daniel Csillag @dccsillag.xyz · 02/09/2024

Useful fact: you can bound the gap between any two probability distributions using f-divergences. A rather nice case is with the Pearson chi-squared f-divergence.

000