Bigbee Labs

In 1969, psychologists Carolyn Kent Rovee and David T. Rovee tied a ribbon from a 10-week-old infant’s ankle to a mobile so that each kick made it move. Within minutes, the babies learned the connection — kicking three times more often than controls. The study, “Conjugate reinforcement of infant exploratory behavior,” was one of the first to show that even very young infants can detect and remember contingencies between their own actions and changes in the world.

I’ve been fretting about Ren and knowledge engineering.

Knowledge engineering traditionally means the process of building and maintaining structured, machine-interpretable knowledge (ontologies, rules, relationships, and representations) usually for reasoning systems.

Ren uses exactly these kinds of things for reasoning. So, it would be natural to set off on a knowledge engineering journey to enable Ren’s reasoning processes. But knowledge engineering, as it turns out, is hard. It’s where symbolic reasoning systems often bog down: creating, curating, and evolving schemas, ontologies, and exemplars at scale. Knowledge engineering is one of the problems that crippled the expert-system era. Not reasoning, but knowledge acquisition.

And all of that sounds, well, bad.

So, I’m thinking about a different way. Something I’m calling “Epistemic Construction”. It’s an idea for how Ren can bootstrap it’s understanding of the world. Sort of like the baby learning about the mobile. Here’s the basics:

1. I have metadata that describes the shape of knowledge. Here’s what a curiosity looks like. Here’s what a inquiry looks like. An event. A condition. A diagnosis. A therapy. A medication. And so on. The metadata describes what Ren needs to know about each term or concept.
2. When Ren is interpreting human language (either in a conversation, or when processing a document like a discharge note), Ren projects the metadata onto the human language – which is Ren’s way of saying, “is there anything in here that looks like a condition, diagnosis, therapy, etc.?”
3. If Ren finds things that look like matches (this discharge order includes a diagnosis), then Ren checks it’s existing knowledge of diagnoses to see if this is something it recognizes. If not, then Ren generates a tentative new bit of knowledge.
4. An expert user, then, can review this new bit of tentative knowledge (along with all the relevant provenance for the tentative knowledge) and confirm that Ren got it right. Or, maybe, correct or adjust the knowledge and then confirm it.
5. Once the human confirms the tentative knowledge, it becomes resolved knowledge and Ren can use it for reasoning purposes.

It’s like Ren is saying:

I saw this new thing. Here’s what I think it is and what it means.

And then a human says:

Yes, Ren. That’s right. (Or provides a correction.)

And then Ren goes about its business, using this new concept.

That’s when Ren discovers the mobile.