PAWS : A novel method of extending distance-metric loss used in self-supervised methods such as BYOL & SwAV to a semi-supervised setting.
Paper Abstract:
Propose PAWS, a novel method of learning, extending the distance-metric loss used in self-supervised methods such as BYOL and SwAV to a semi-supervised setting
+ Set new state-of-the-art for ResNet-50 on ImageNet trained with either 10% or 1% of the labels, reaching 75% and 66% top-1 respectively (achieved with 4x — 12x less training)
+ Match performance of fully supervised learning with bigger networks, while using 10x fewer labels
+ All code available on GitHub! [ Ссылка ]
Speaker Bio:
Mido is a researcher at Facebook AI Research (FAIR) and Mila – Quebec AI Institute.
+ He is an NSERC Vanier Scholar and holds a Vadasz Doctoral Fellowship in Engineering at McGill University.
+ He is particularly interested in the capability of computers to efficiently develop a perceptual understanding of our world from limited human supervision. His research has been featured in several media outlets, including VentureBeat, TechCrunch, and SiliconANGLE.
+ He has also served as a reviewer for numerous machine learning and control conferences and journals, most recently as an expert reviewer for ICML.
Social links:
Twitter handle: mido_assran
Github handle: MidoAssran
