Hi everyone,
I’m trying to clarify a very practical point in Monty’s architecture.
Before posting here I’ve read the forum discussion Some thoughts about scale invariance and model composition and watched Jeff Hawkins’ talk on scale invariance in the brain (https://www.youtube.com/watch?v=gt_M7z8UkiM). These post and video overlap with what I’m asking, but I don’t see a direct answer there. So, let me put it here.
According to the documentation, each Learning Module (LM) can store an object model only within a fixed max_size (for example, 10 cm × 10 cm × 10 cm). If a large portion of observations falls outside this grid, the LM either stores only a small local fragment or refuses to form a single object model at all.
Now consider a simple example.
An agent is far away and observes the Earth from space. For a large-scale LM, the entire planet fits within its max_size, so it can form an object model like “planet”.
Later, the agent moves closer and starts exploring the surface. Smaller-scale LMs now form object models of local things: rocks, trees, buildings, terrain patches, etc.
My question is very simple:
How does Monty understand that the “planet Earth” seen from far away and all these local objects seen up close belong to the same physical object?
In other words:
Is there a mechanism in Monty that explicitly links an original large-scale object model with many smaller-scale object models as parts of the same thing?
Or do different LMs always operate on independent objects, with no explicit representation of such cross-scale unity?
I’m trying to understand whether this linkage is already implicit in the current Monty architecture (e.g. via voting, memory, or hypothesis passing), or whether it is an open problem / future work beyond the current design.
