Use Cases

🚧

This doc is out of date and pending removal.

See https://github.com/diffgram/diffgram

Create, Update, And Maintain Datasets

• Create, Update, and Maintain a single Dataset (optional train/val/test splits).
• Create, Update, and Maintain many Datasets.
• Create, Update, and Maintain many Datasets with interrelationships.

Create processes for working with Deep Learning systems

• Reduce single points of failure (single data scientist, single annotator, single computer node)
• Distribute some human control to many annotation firms (including those by API)
• Create a system of record & version control

Compliance and threat actors

• Defeat adversarial approaches by rapidly retraining
• Compliance (ie Who labeled it? Who can export sets?)

Launch faster

• Faster time to market with automation and ready to go software
• Faster time to market with faster hypothesis testing and iteration
• Faster time to market with rapid response to changing distributions

Control costs

• Reduce engineering effort (ie Integration)
• Monitor annotation costs, Monitor permissions Monitor overall workflow between including at boundaries (Kanban board from import through model triggers / return)
• Speed up annotation (curation, literal annotation speedups)

Reduce engineering burden

• For literal connection from data to ml hardware
• In the Training Data process itself ie with Prep, Annotation, Datasets
• Reduce Total Cost of Ownership (ToC),
  ie formats, integrations, new techniques, reviews etc…
• Monitor dataset effectiveness

Explore more

• Explore and iterate on ideas for how to structure Datasets.
• Create better datasets by involving Subject Matter Experts earlier and more often