monthly status 20220203 en mid.txt

Monthly Status - Secure AI Fabric - January 2022

The Enterprise Neurosystem Secure AI Fabric track started meeting in January. The purpose of this track is to work on how an AI system can access and manage data originating in a secure environment, subject to governance constraints, access/security issues, and multiparty sharing considerations. Integration, middleware, messaging and networking technologies are discussed.

This month, we started to explore the following topics.

• Security of models

  Governance constraints for the industries mostly represented in the group require us to ask questions about data that have not been adequately addressed in the industry currently. For example, distinction between data in-flight vs. at-rest is very important in FSI and Telco (in-flight is OK for public cloud, at-rest is not). This requires thinking beyond existing data lake methodology, which we are exploring.

• Shared catalog

  One of the key use cases is to leverage someone else's hardware for training. This means that the model catalog needs to be shared, and therefore the format of the shared catalog must be agreed. IBM Research has contributed a number of resources to help develop this part of the project. The idea is that it will be more of a "digital asset" catalog that will accommodate multiple types of model formats, as well as any supporting artifacts required for training.

• Training / Job scheduling

  We have begun a discussion of how certain types of models will be mapped to specific hardware that can execute those models. This goes beyond standard notions within Kubernetes, because it requires knowledge of different types of hardware, such as Graphcore, Lambda, etc. This is very important in the research / DoE use cases. Ultimately, this could even become an optimization problem, trading off cost for time to completion.

• Plans for next month

  During February, we look forward to progress from the team that IBM Research has kindly donated, to write the shared catalog code. We will also hold exploration demos to understand how to fill the other gaps with existing open source technologies. Finally, we plan to have relevant research presentations if they are pertinent to the topics of the track.