Great to see the concreteness of this example, some thoughts on the candidate properties:
The relationships between latent variables can change under ontology shifts, but we still want the semantics of our latent variables to remain invariant in some sense. This means we need some other variables tracking the relationships between our latent variables, & we’d update on those variables during ontology shifts. Latent variables are partly selected on the criterion that they can adapt to these changes in relationships while still maintaining the correspondence principle.
There’s a problem of whether we want to think of factorization/latent variables as a stance/frame we’re taking towards the system or an objective structural property of the system. I lean towards the former because when humans model the minds of other humans, they think of those minds as making use of similar abstractions to think & reason despite almost never observing the actual internals of human brains. The degree of freedom in the “latent variable embedding” of nuclear exchange initiation also suggests some notion of subjectivity.
The fact that we have to identify nuclear exchange initiation across a wide variety of scenarios points to ‘re-use’ being an important basis of convergent factorization. Loopiness comes in when our existing reusable abstractions affect what new abstractions/models of the world we’re able to construct.
Great to see the concreteness of this example, some thoughts on the candidate properties:
The relationships between latent variables can change under ontology shifts, but we still want the semantics of our latent variables to remain invariant in some sense. This means we need some other variables tracking the relationships between our latent variables, & we’d update on those variables during ontology shifts. Latent variables are partly selected on the criterion that they can adapt to these changes in relationships while still maintaining the correspondence principle.
There’s a problem of whether we want to think of factorization/latent variables as a stance/frame we’re taking towards the system or an objective structural property of the system. I lean towards the former because when humans model the minds of other humans, they think of those minds as making use of similar abstractions to think & reason despite almost never observing the actual internals of human brains. The degree of freedom in the “latent variable embedding” of nuclear exchange initiation also suggests some notion of subjectivity.
The fact that we have to identify nuclear exchange initiation across a wide variety of scenarios points to ‘re-use’ being an important basis of convergent factorization. Loopiness comes in when our existing reusable abstractions affect what new abstractions/models of the world we’re able to construct.