Towards the Generalization of Contrastive Self-Supervised Learning

Authors: Weiran Huang, Mingyang Yi, Xuyang Zhao, Zihao Jiang

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct various experiments on the real-world datasets and observe that the downstream performance of contrastive SSL is highly correlated to the concentration of augmented data in Section 5. Our experiments are conducted on CIFAR-10 and CIFAR-100 (Krizhevsky, 2009).
Researcher Affiliation Collaboration Weiran Huang1 Mingyang Yi2 Xuyang Zhao3 Zihao Jiang1 1 Qing Yuan Research Institute, Shanghai Jiao Tong University 2 Huawei Noah s Ark Lab 3 School of Mathematical Sciences, Peking University
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code No No explicit statement about releasing source code or link to a code repository is provided.
Open Datasets Yes Our experiments are conducted on CIFAR-10 and CIFAR-100 (Krizhevsky, 2009).
Dataset Splits No Our experiments are conducted on CIFAR-10 and CIFAR-100 (Krizhevsky, 2009). ... Each model is trained with a batch size of 512 and 800 epochs. To evaluate the quality of the encoder, we follow the KNN evaluation protocol (Wu et al., 2018). No explicit mention of train/validation/test dataset splits.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are provided for the experimental setup.
Software Dependencies No The paper mentions algorithms and models like "ResNet-18" and "Sim CLR", but does not specify software dependencies with version numbers (e.g., PyTorch 1.9, CUDA 11.1).
Experiment Setup Yes We use Res Net-18 (He et al., 2016) as the encoder, and the other settings such as projection head remain the same as the original settings of algorithms. Each model is trained with a batch size of 512 and 800 epochs. We compose all 5 kinds of transformations together, and then successively drop one of the composed transformations from (e) to (b) to conduct 5 experiments for each dataset (Table 1).