Here is a quick video where I talk a little bit about an AI application that has bounced around my head for years: the Perception Engine (PE). It’s basically a high-fidelity recreation of the world as it’s sensing it, but slightly lower fidelity than our open world. It allows for preprocessing of the environment before the AI gets any input. This allows the AI to be decoupled from it. Action would take place in the real world, of course, and their consequences will then be reflected in the PE. Cooperating AIs could share the same PE.
Benefits include the AI being able to overlay or augment what it sees to have an ‘imagination’ so it can simulate potential actions in real-time before actually taking that action. It would also have basic object recognition done so that the AI can focus on other things. It would allow symbol grounding for any symbolic processing from the AI. The PE could be improved separately from the AI.
There are some potential detriments to using a PE for an AI. Processing time may delay PE updates, leading to inaccurate input. Any bad preprocessing can still lead to garbage data for the AI. The storage required for even a simple PE could be large, depending on how long the data is kept around.
This is a project that I’ll be taking on during my Master’s in AI, which I start this fall. More on that very soon.