Saddle-to-Saddle Dynamics in Diagonal Linear Networks

Authors: Scott Pesme, Nicolas Flammarion

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide numerical experiments to support our findings.
Researcher Affiliation Academia Scott Pesme EPFL scott.pesme@epfl.chNicolas Flammarion EPFL nicolas.flammarion@epfl.ch
Pseudocode Yes Algorithm 1: Successive saddles and jump times of limα 0 βα
Open Source Code No The paper does not provide any explicit statements about releasing open-source code or links to a code repository.
Open Datasets No For each experiment we generate our dataset as yi = xi, β where xi = N(0, H) for a a diagonal covariance matrix H and β is a vector of Rd. The only assumption we make on the data throughout the paper is that the inputs (x1, . . . , xn) are in general position.
Dataset Splits No The paper describes how the data is generated for each experiment, but does not specify training, validation, or test splits. It directly uses the generated data for the numerical experiments.
Hardware Specification No The paper describes the experimental setup, but does not specify any hardware details (e.g., CPU, GPU, memory) used for running the experiments.
Software Dependencies No The paper mentions that "Gradient descent is run with a small step size" but does not provide specific software names with version numbers for reproducibility.
Experiment Setup Yes For each experiment we generate our dataset as yi = xi, β where xi = N(0, H) for a a diagonal covariance matrix H and β is a vector of Rd. Gradient descent is run with a small step size and from initialisation u0 = √2α1d and v0 = 0d for some initialisation scale α > 0. Figure 1 and Figure 4 (Left): (n, d, α) = (5, 7, 10−120), H = Id, β = (10, 20, 0, 0, 0, 0, 0) ∈ R7.