Closed XFeiF closed 3 years ago
Part 1: Introduction: Unsupervised learning from vision
Part 2: Representation learning from images and videos
Part 3: Unsupervised learning from 3D
Part 4: Self-supervised learning for Robotics
Unsupervised visual learning usefulness.
Learning from videos (sequentially correlated data) or images (spatially correlated data) provides a lot of information for feature representation!
Learning from 3D - very valuable as it provides structure (and constraints) for learning.
Learning with self-supervision and interactions - provides more task-specific "supervision"; can directly learning outcomes; provides better features for subsequent tasks.
Without labels, how do we learn representations?
We can learn from context, color, video, sound and invariances.
[part1] [part2]