Benign overfitting in leaky ReLU networks with moderate input dimension

Authors: Kedar Karhadkar, Erin George, Michael Murray, Guido F. Montufar, Deanna Needell

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To further support our theory, we train shallow neural networks on the data model described in Definition 2.1 and record the numerical results. Scripts to reproduce these experiments can be found at https://github.com/kedar2/benign_overfitting. These experiments were run on the CPU of a Mac Book Pro M2 with 8GB of RAM.
Researcher Affiliation Academia Kedar Karhadkar1 Erin George1 Michael Murray1 Guido Montúfar12 Deanna Needell1 {kedar,egeo,mmurray,montufar,deanna}@math.ucla.edu 1UCLA 2Max Planck Institute for Mathematics in the Sciences Equal contribution
Pseudocode No The paper contains mathematical equations and descriptions of algorithms, but no structured pseudocode blocks or algorithm boxes are presented.
Open Source Code Yes Scripts to reproduce these experiments can be found at https://github.com/kedar2/benign_overfitting.
Open Datasets No 2.1 Data model We study data generated as per the following data model. Definition 2.1. Suppose d, n, k N, (0, 1) and v Sd 1. If (X, y, y, x, y) D(d, n, k, , v) then 1. X Rn d is a random matrix whose rows, which we denote xi, satisfy xi = yiv + 1 ni, where ni N(0d, 1 d(Id vv T )) are mutually i.i.d..
Dataset Splits Yes Parameter settings: = 0.1, = 5/n, m = 64, k = 0.1n, number of trials = 5, size of validation sample = 1000.
Hardware Specification Yes These experiments were run on the CPU of a Mac Book Pro M2 with 8GB of RAM.
Software Dependencies No The paper describes the learning algorithm (gradient descent with hinge loss) but does not specify any particular software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow).
Experiment Setup Yes Parameter settings: = 0.1, = 5/n, m = 64, k = 0.1n, number of trials = 5, size of validation sample = 1000.