notes 20220118 en mid.txt

Secure AI Fabric - 1/18/2022

Shared Catalog Discussion

• Agreeing on model format
• Model self-describing
  • Link to Data
  • Link to Metadata
  • OMX format, pytorch format, etc [ryan -- big fan of h5]
  • Data - list of models it has been used to train
  • Model - list of data sets it has been used on
  • Shared catalog can be not just an AI model catalog, but a "digital asset" catalog
  • Title of slide - redo to discuss "assets" rather than "models"
    • Access to models rather than models

Training

• "Heat maps" of data and hardware
• Rules as to whether it is better to move data to hardware or not
• Certain types of hardware are only suited to certain types of problem
  • Can be matched via metadata
• There will always be a Venn diagram as per how something can be executed
• DoE has said that AI will be very different from traditional stovepipe approaches

Scheduling

• Assume that we are using containers
• Specialized hardware (Graphcore) issue
  • Need to have a discussion with them and explain the opportunity
  • Worst case, for corner cases, would need proxy container