Incremental Learning in Diagonal Linear Networks

Authors: Raphaël Berthier

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical In this paper, we describe the trajectory of the gradient flow of DLNs in the limit of small initialization. We show that incremental learning is effectively performed in the limit: coordinates are successively activated, while the iterate is the minimizer of the loss constrained to have support on the active coordinates only. This shows that the sparse implicit regularization of DLNs decreases with time. This work is restricted to the underparametrized regime with anti-correlated features for technical reasons.
Researcher Affiliation Academia Rapha el Berthier EMAIL EPFL Lausanne, Switzerland
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. It focuses on mathematical proofs and theoretical analysis.
Open Source Code No The paper does not contain any explicit statements about code availability, specific repository links, or mention of code in supplementary materials.
Open Datasets No In this simulation, n = 3 and the data X Rn d, y Rn is generated randomly with i.i.d. standard Gaussian entries, conditionally on the event that Assumptions (A1) and (A2) hold.
Dataset Splits No The paper describes synthetic data generated randomly for simulations to illustrate theoretical concepts, rather than using pre-existing datasets with defined splits for training, validation, or testing.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running its simulations or computations.
Software Dependencies No The paper does not mention any specific software dependencies or libraries with version numbers (e.g., Python, PyTorch, TensorFlow) that would be needed to replicate the work.
Experiment Setup Yes The simulations are run with ε = 10 8 (left plots) and ε = 10 20 (right plots). In this simulation, n = 5, d = 4 and the data X Rn d, y Rn is generated randomly with i.i.d. standard Gaussian entries, conditionally on the event that Assumptions (A1) and (A2) hold. The initialization is θ(ε)(0) = (ε, . . . , ε) and thus k = (1, . . . , 1).