The Power of Random Features and the Limits of Distribution-Free Gradient Descent

Authors: Ari Karchmer, Eran Malach

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical Our main result shows that if a parametric model can be learned using mini-batch stochastic gradient descent (b SGD) without making assumptions about the data distribution, then with high probability, the target function can also be approximated using a polynomialsized combination of random features. Along the way, we also introduce a new theoretical framework called average probabilistic dimension complexity (adc), which extends the probabilistic dimension complexity developed by Kamath et al. (2020). We prove that adc has a polynomial relationship with statistical query dimension, and use this relationship to demonstrate an infinite separation between adc and standard dimension complexity.
Researcher Affiliation Academia 1Harvard University, Cambridge, Massachusetts, USA.
Pseudocode Yes Algorithm 1 Predict(z) 1: input: point z; example oracle access to function f µ, with respect to example distribution ρ. 2: Sample g µ. 3: Sample x, f(x) from example oracle. 4: predict: g(z) g(x) f(x)
Open Source Code No The paper does not contain any explicit statements or links indicating the release of open-source code for the methodology described.
Open Datasets No This paper is theoretical and focuses on mathematical proofs and definitions; it does not present empirical results or use specific datasets that would be made publicly available.
Dataset Splits No This is a theoretical paper that does not involve empirical experiments requiring dataset splits.
Hardware Specification No This is a theoretical paper and does not describe any experimental hardware used.
Software Dependencies No This is a theoretical paper and does not specify any software dependencies with version numbers.
Experiment Setup No This is a theoretical paper and does not include details on experimental setup, hyperparameters, or training configurations.