Learning Optimal Linear Regularizers

Authors: Matthew Streeter

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now evaluate the Learn Lin Reg and Tune Reg algorithms experimentally, using both real and synthetic data.
Researcher Affiliation Industry Matthew Streeter 1 1Google Research. Correspondence to: Matthew Streeter <mstreeter@google.com>.
Pseudocode Yes Algorithm Learn Lin Reg Input: Set of (validation loss, training loss, feature vector) tuples n (Vi, ˆLi, qi) | 1 i m o . Sort tuples in ascending order of Vi, and reindex so that V1 V2 . . . Vm. for i from 1 to m do
Open Source Code Yes Code for both algorithms is available on Git Hub (Streeter, 2019).
Open Datasets Yes We then consider two problems that make use of deep networks trained on MNIST and Image Net (Russakovsky et al., 2015).
Dataset Splits No The paper mentions using "validation loss" and describes the "flowers" dataset as "split evenly into training and test," but it does not provide specific training/validation/test dataset splits (percentages or sample counts) for all datasets used (e.g., MNIST, ImageNet) that would be needed for full reproducibility of data partitioning.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper refers to "Tensor Flow Team" but does not specify any software dependencies (e.g., libraries, frameworks, or solvers) with explicit version numbers.
Experiment Setup Yes For the softmax regression problems, we obtain a nearoptimal solution by running Ada Grad for 100 epochs. Each grid Λi contains 50 points. For L1 and L2, the grid is loguniformly spaced over [.1, 100], while for label smoothing and dropout it is uniform over [0, 1].