Taskonomy - Disentangling Task Transfer Learning

Explore the concept of task transfer learning in computer vision through this 31-minute lecture from the University of Central Florida. Delve into topics such as transfer modeling, higher-order transfers, and normalization techniques. Learn about computing global taxonomies and examine experiments that test the sanity of trained task-specific networks. Evaluate computed taxonomies and their generalization to novel tasks. Investigate the significance testing of structures and assess the findings' applicability across different datasets, including MIT Places and ImageNet. Gain insights into task similarity trees and their implications for efficient transfer learning in various computer vision applications.

Transfer Modeling
Higher Order Transfers
Normalization
Computing Global Taxonomy
Experiments
Sanity of trained Task Specific Networks
Evaluation of Computed Taxonomies
Generalization to novel tasks
Significance Test of the Structure
Evaluation on MIT Places & ImageNet Are the findings dataset dependent?
Task Similarity Tree