Deep Learning meets Projective Clustering

Authors: Alaa Maalouf, Harry Lang, Daniela Rus, Dan Feldman

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on the GLUE benchmark yield networks that are both more accurate and smaller compared to the standard matrix factorization (SVD).
Researcher Affiliation Academia 1 Robotics & Big Data Labs, Department of Computer Science, University of Haifa 2 CSAIL, MIT
Pseudocode No The paper describes a pipeline with numbered steps but does not present it as formal pseudocode or an algorithm block.
Open Source Code Yes Complete open code to reproduce the resulting networks is provided.
Open Datasets Yes GLUE benchmark. We run our experiments on the General Language Understanding Evaluation (GLUE) benchmark (Wang et al., 2018).
Dataset Splits No The paper mentions fine-tuning for '2 epochs' and evaluating on the 'training set' for model selection in Appendix D, but does not specify explicit training/validation/test dataset splits with percentages or counts for the general experiments.
Hardware Specification Yes All the experiments were conducted on a AWS c5a.16xlarge machine with 64 CPUs and 128 RAM [Gi B].
Software Dependencies Yes Transformers version 3.1.0, and Py Torch version 1.6.0 (Paszke et al., 2017).
Experiment Setup Yes Given an embedding layer from a network that is trained on a task from GLUE, an integer k 1, and an integer j 1. We build and initialize a new architecture that replaces the original embedding layer by two smaller layers as explained in Figure 3. We then fine tune the resulted network for 2 epochs.