Understanding Augmentation-based Self-Supervised Representation Learning via RKHS Approximation and Regression

Authors: Runtian Zhai, Bingbin Liu, Andrej Risteski, J Zico Kolter, Pradeep Kumar Ravikumar

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we demonstrate this point with mask-type augmentations on synthetic and real datasets, and show that (i) κ depends on both the augmentation strength and the augmentation strategy; (ii) a smaller κ (e.g. stronger augmentation) leads to a smaller generalization gap, but an overly strong augmentation causes poor training performance. Thus, there is a sweet spot in the middle with the best test performance.
Researcher Affiliation Academia Runtian Zhai, Bingbin Liu, Andrej Risteski, Zico Kolter, Pradeep Ravikumar Carnegie Mellon University {rzhai,bingbinl,aristesk,zkolter,pradeepr}@cs.cmu.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code of Section 5 can be found at https://colab.research.google.com/drive/1lo SZLLI-qfo KE7BCIi1SWJKgru U6i4ku?usp=sharing.
Open Datasets Yes We demonstrate on the NLP dataset wikipedia-simple. We study masked language modeling, where x is a full sentence and a is a masked sentence. using QNLI (Wang et al., 2018) and SST-2 (Socher et al., 2013).
Dataset Splits No The paper uses QNLI and SST-2 datasets but does not explicitly state the training, validation, and test splits used for these datasets.
Hardware Specification Yes We use 8 NVIDIA A6000 GPUs for pretraining. We use 4 NVIDIA A6000 GPUs for downstream training and evaluation.
Software Dependencies No The paper mentions using "roberta-large models" and the "Huggingface official repository" but does not provide specific version numbers for software dependencies such as libraries, frameworks, or programming languages.
Experiment Setup Yes The classifiers are trained for 3 epochs on QNLI and 6 epochs on SST-2.