Method
Taskonomy (Task Taxonomy) Graph
Directed hypergraph that captures the notion of task transferability over any given task dictionary.
- nodes: tasks (source(s) and target)
- edge: feasible transfer case.
- edge weight: prediction of performance of target given sources
Taxonomy Computation
- produces a family of taskonomy graphs, parameterized by the available supervision budget, chosen tasks, transfer orders, and transfer functions’ expressiveness.
- 4 step process
Step 1: Task Specific Modeling
Train task specific networks for each source task.
Step 2: Transfer Modeling
All feasible transfers between sources and targets are trained (including higher order transfers ie multiple input tasks to train one target).
Step 3: Task Affinity Normalization
Task affinities acquired from transfer function performances are normalized.
- basically takes all the losses that you compute for every source, target pair and normalizes them so you can compare them (ie they're in the same space)
Step 4: Compute Taxonomy
Description
Overview
