Autonomous Adaptive Edge-Distributed ML

This project will consider the problem of updating and adapting locally deployed ML models within edge-agents based on local knowledge, or incoming information. We will set up an overall framework of information and adaptation.

There are three key aspects to this work: 1) How to adapt agents to retain existing knowledge, but include further information.

How to know when new knowledge changes/and perturbs a model for a changing system, and hence requires some level of forgetting or just enhances knowledge about a more static one. This will be the initial consideration.

We will look at employing random network distillation and zero-shot model distillation as a proxy for assessing information that must be retained. 2) How model adaptation can be done on a constrained device in an efficient way. 3) A distributed value-of-information model that can assess when information is of additional value, that can choose to obtain information, and knowing when processing is worth the cost.

This leads to an overall information-transactional model where agents can transact with other agents to obtain information.