[FEATURE] Training Curriculum

[TSampley]

User Story

As a Model Trainer, I want to create structured learning curriculums, so that I can run training epochs on increasingly complex data and reduce training time while increasing fidelity of latent space

Acceptance Criteria

• Given an agent role; knowledge expected of that role
  • When I request a training graph for the given role
    • Then I receive a graph describing a training process for a model
    • And each node represents a "subject" - a set of data associated with a single topic
    • And each subject is linked to its prerequisites
    • And each subject includes related performance benchmarks

References

• training a network on a smaller data set is less expensive
• knowledge and concepts are cumulative, i.e. they build on prior knowledge and concepts
• training a network on a complete dataset means the latent space is influenced by the entire dataset. Weighting can influence the "strength" of each contribution, but experience has shown that models can fail to regress in productive ways and training needs to be restarted from an earlier stage or even from scratch.