Benchmarking Self-Supervised Learning on STL-10: SimCLR Vs BYOL

Abstract

Self-supervised learning (SSL) has emerged as an effective paradigm for learning visual representations without reliance on labeled data. This study presents a controlled benchmark of two widely adopted SSL methods, SimCLR and BYOL, evaluated on the STL-10 dataset. Both methods are implemented using an identical ResNet-18 backbone and trained under matched computational and optimization settings. Representation quality is assessed through linear probing and k-NN classification. Under these constraints, SimCLR demonstrates stronger performance than BYOL, achieving a linear probe accuracy of 71.21% compared to 69.90% for BYOL. These results emphasize practical considerations in SSL benchmarking and highlight performance trade-offs that arise under resource-limited training regimes.