SGD Learns Over-parameterized Networks that Provably Generalize on Linearly Separable Data

Authors: Alon Brutzkus, Amir Globerson, Eran Malach, Shai Shalev-Shwartz

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Figure 1 we demonstrate this empirically for a linearly separable data set (from a subset of MNIST) learned using over-parameterized networks.
Researcher Affiliation Academia Alon Brutzkus & Amir Globerson The Blavatnik School of Computer Science Tel Aviv University, Israel alonbrutzkus@mail.tau.ac.il,amir.globerson@gmail.com Eran Malach & Shai Shalev-Shwartz School of Computer Science The Hebrew University, Israel eran.malach@mail.huji.ac.il,shais@cs.huji.ac.il
Pseudocode No The paper describes the SGD update rule mathematically (Eq. 3) but does not provide structured pseudocode or an algorithm block.
Open Source Code No The paper does not contain any statement about releasing open-source code or a link to a code repository.
Open Datasets No The linearly separable data set consists of 4000 MNIST images with digits 3 and 5, each of dimension 784.
Dataset Splits No The size of the training set is 3000 and the remaining 1000 points form the test set.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes The setting of Section 5 is implemented (e.g., SGD with batch of size 1, only first layer is trained, Leaky Re LU activations) and SGD is initialized according to the initialization defined in Eq. 6.