Explore self-supervised representation learning and contrastive techniques in computer vision through this comprehensive 58-minute lecture by Stanford University PhD student Nandita Bhaskhar. Dive deep into six recent frameworks: SimCLR, MoCo V2, BYOL, SwAV, DINO, and Barlow Twins. Examine their methodologies, performance, strengths, and weaknesses, with a focus on potential applications in the medical domain. Gain insights into how these techniques can leverage unlabeled datasets, overcoming the limitations of traditional supervised learning approaches. Learn about the speaker's research on observational supervision, self-supervision for medical data, and out-of-distribution detection for clinical deployment. Benefit from a thorough exploration of topics including invariant representations, pre-text tasks, entity discrimination, and various architectural approaches in self-supervised learning.
Overview
Syllabus
Intro
SS Learning: Invariant Representations
Pre-text Tasks: A Deeper Dive
Contrastive Learning: Entity Discrimination
Contrastive Learning: Problem
SimCLR: Simple Contrastive Learning Representatio
SimCLR: Architecture
SimCLR: Loss Function
SimCLR: Findings
MoCo V2: Momentum Contrast
MoCo V2: Architecture
MoCo V2: Main Principle
MoCoV2: Loss Function
MoCo V2: Findings
BYOL: Bootstrap Your Own Latent
BYOL: Architecture
BYOL: Main Principle
BYOL: Findings
SWAV: Swapping Assignments between Views
SWAV: Architecture
SWAV: Loss Function
SWAV: Main Principle
SWAV: Multi-crop
SWAV: Additional Findings
DINO: Self-Distillation with NO labels
DINO: Attention-Maps
VIT (Vision Transformer): Architecture
DINO: Architecture
DINO: Loss Function
DINO: Main Principle
DINO: Multi-crop
DINO: Additional Findings Compute
Taught by
Stanford MedAI
