Understanding Deep Architecture with Reasoning Layer

Authors: Xinshi Chen, Yufei Zhang, Christoph Reisinger, Le Song

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments aim to validate our theoretical prediction with computational simulations, rather than obtaining state-of-the-art results. We conduct two sets of experiments, where the first set of experiments strictly follows the problem setting described in Sec 2 and the second is conducted on BSD500 dataset [56] to demonstrate the possibility of generalizing the theorem to more realistic applications. Implementations in Python are released.
Researcher Affiliation Academia Xinshi Chen Georgia Institute of Technology xinshi.chen@gatech.edu Yufei Zhang University of Oxford yufei.zhang@maths.ox.ac.uk Christoph Reisinger University of Oxford christoph.reisinger@maths.ox.ac.uk Le Song Georgia Institute of Technology lsong@cc.gatech.edu
Pseudocode No The paper presents algorithm update steps (e.g., for GD and NAG) but does not provide a formally structured pseudocode block or algorithm box.
Open Source Code Yes Implementations in Python are released1. 1https://github.com/xinshi-chen/Deep-Architecture-With-Reasoning-Layer
Open Datasets Yes We split BSD500 (400 images) into a training set (100 images) and a test set (300 images).
Dataset Splits No The paper states the training and test set sizes for the BSD500 dataset, but it does not explicitly mention a separate validation set or its split. For synthetic experiments, it states "During training, n samples are randomly drawn from these 10000 data points as the training set" without specifying a validation split.
Hardware Specification No The paper does not specify any hardware details (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper states "Implementations in Python are released" but does not specify the version of Python or any other software libraries used, nor their versions.
Experiment Setup Yes Each model is trained by ADAM and SGD with learning rate grid-searched from [1e-2,5e-3,1e-3,5e-4,1e-4], and only the best result is reported.