Keep the Momentum: Conservation Laws beyond Euclidean Gradient Flows

Authors: Sibylle Marcotte, Rémi Gribonval, Gabriel Peyré

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

Reproducibility Variable Result LLM Response
Research Type Experimental We consider here a 3-layer MLP trained for classification on the MNIST dataset (Le Cun et al., 2010) with the cross entropy loss function and a ReLU non-linearity... Figure 1, left, shows the evolution of the loss for a range of step size δ up to almost no convergence... Figure 2 shows how the evolution of the loss and the preserved quantities for GF is impacted by the momentum parameter µ = 1/τ.
Researcher Affiliation Academia Sibylle Marcotte 1 Rémi Gribonval 2 Gabriel Peyré 1 3 1ENS PSL Univ. 2Univ Lyon, Ens L, UCBL, CNRS, Inria, LIP. 3CNRS. Correspondence to: Sibylle Marcotte <sibylle.marcotte@ens.fr>.
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our code to compute them is available at https://github.com/sibyllema/Conservation_laws_ICML.
Open Datasets Yes We consider here a 3-layer MLP trained for classification on the MNIST dataset (Le Cun et al., 2010).
Dataset Splits No The paper mentions using training and test sets from MNIST but does not specify a validation set or explicit split percentages for training, validation, and testing.
Hardware Specification No The paper does not specify the hardware used for running the experiments, such as specific GPU or CPU models.
Software Dependencies No We used the software Sage Math (The Sage Developers, 2022), which relies on a Python interface.
Experiment Setup Yes We consider the following time discretization of the flows, where time at step k is t = kδ and δ > 0 is the time step... This can be re-written in the usual form of a gradient descent with momentum θk+1 = θk αMk EZ(θk) + β(θk θk 1) where α := δ ν + µ/δ and β := µ δν + µ < 1. Here β [0, 1) is the momentum (extrapolation) parameter, so that β = 0 corresponds to usual gradient descent, and setting β = 1 is maximum momentum (which is not in general ensured to converge). Figure 1, left, shows the evolution of the loss for a range of step size δ... Figure 2 shows how the evolution of the loss and the preserved quantities for GF is impacted by the momentum parameter µ = 1/τ.