Self-Supervised Relational Reasoning for Representation Learning

Authors: Massimiliano Patacchiola, Amos J. Storkey

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed method following a rigorous experimental procedure, using standard datasets, protocols, and backbones. Self-supervised relational reasoning outperforms the best competitor in all conditions by an average 14% in accuracy, and the most recent state-of-the-art model by 3%.
Researcher Affiliation Academia Massimiliano Patacchiola School of Informatics University of Edinburgh mpatacch@ed.ac.uk Amos Storkey School of Informatics University of Edinburgh a.storkey@ed.ac.uk
Pseudocode Yes An overview of the model is given in Figure 1 and the pseudo-code in Appendix C (supp. material).
Open Source Code Yes code released with an open-source license1; 1https://github.com/mpatacchiola/self-supervised-relational-reasoning
Open Datasets Yes using standard datasets (CIFAR-10, CIFAR-100, CIFAR-100-20, STL-10, tiny-Image Net, Slimage Net)
Dataset Splits Yes We follow the linear evaluation protocol defined by Kolesnikov et al. (2019) training the backbone for 200 epochs using the unlabeled training set, and then training for 100 epochs a linear classifier on top of the backbone features (without backpropagation in the backbone weights).
Hardware Specification No The paper does not specify the hardware used for the experiments (e.g., GPU models, CPU types).
Software Dependencies No We provide a standardized environment implemented in Pytorch (see code in supp. material). (Only Pytorch is mentioned, without a version number, and no other software dependencies with versions are specified.)
Experiment Setup Yes Hyperparameters (relation learner): mini-batch of 64 images (K = 16 for Res Net32 on tiny-Image Net, K = 25 for Res Net-34 on STL-10, K = 32 for the rest), Adam optimizer with learning rate 10 3, binary cross-entropy loss with focal factor (γ = 2). Relation module: MLP with 256 hidden units (batch-norm + leaky-Re LU) and a single output unit (sigmoid). Aggregation: we used concatenation as it showed to be more effective (see Appenidx B.8, Table 13 supp. material). Augmentations: horizontal flip (50% chance), random crop-resize, conversion to grayscale (20% chance), and color jitter (80% chance). Backbones: Conv-4, Res Net-8/32/56 and Res Net-34 (He et al., 2016).