Question about continuity and identity in distributed models

Research and Theory
, ,
1

Hi everyone,

I’ve been exploring the Thousand Brains framework recently, and I’m trying to better understand how to think about system-level continuity over time in a model composed of many interacting learning modules.

My background is in mechanistic modeling and complex systems (physics + biology + AI), so I may be approaching this from a different angle — apologies if I’m missing something obvious.

From what I understand, each learning module builds its own model of objects using sensorimotor interaction and reference frames, and these modules interact through voting and messaging to reach consensus about the world.

This makes a lot of sense to me at the level of inference.

However, I’m trying to understand something slightly different:

:backhand_index_pointing_right: How should we think about the persistence of the system as a coherent entity over time, while all of its internal models are continuously being updated?

In other words:

  • What defines the identity of the system across time?
  • How is stability of behavior maintained while learning is ongoing?
  • Is there a notion of direction or continuity beyond moment-to-moment consensus?

This question became relevant to me while working on behavior and neural systems at the Max Planck Institute for Neurobiology of Behavior, where I became interested in how coherent behavior can emerge from distributed components without central control.

More recently, I explored similar questions in an agent-based system at the Santa Fe Institute, where LLM-driven agents interacted in space and time. In that context, coherence didn’t come just from agreement, but from continuous interaction — agents maintained some consistency in behavior while still adapting.

Trying to frame this more precisely, I’ve been thinking about three aspects that might complement the current picture:

  • Self-maintenance
    Systems that actively preserve their internal organization while interacting
    (in the sense of autopoiesis, e.g. Maturana & Varela)

  • Structural coupling
    Systems that co-evolve with their environment rather than passively representing it

  • Emergent directionality
    Behavior that appears goal-directed without an explicit external objective, but arises from the dynamics of the system

I’m wondering whether something along these lines is already considered within the Thousand Brains framework, or if the current focus is mainly on representation and inference.

I might be mixing concepts here, but I’d be very interested to hear how people think about this.

Thanks!
Emanuel

3 Likes
2

Hi @Emanuel_Cura_Costa and thank you for the great and detailed question!

Sorry about the late reply, we had some deadlines and a focus week, and your question is not a quick one to answer.

I hope I am not misunderstanding some of the terms you use, as they seem like they probably have very specific definitions in your field. I’ll try to give some general context, and then maybe you can ask some follow-up questions and steer this conversation in the right direction.

I’ll start with "How should we think about the persistence of the system as a coherent entity over time, while all of its internal models are continuously being updated?"

This is a great question, and it is not fully addressed in Monty since so far, we’ve been working with pretty small amounts of learning modules, no hierarchy, and short time scales. But the general idea is that in order to have stability in the system, different learning modules (or cortical columns) would learn at different speeds. So, for example, an LM at the lowest level of the hierarchy (like a column in V1) should have quite stable models that are only updated slowly. This is important since other LMs rely on that LM’s output, and if the output is constantly changing what it represents, this will be impossible. At higher levels in the hierarchy (in the hippocampus at its most extreme), we could have much faster learning, where we can quickly learn temporary arrangements of objects, like objects are on your table right now. Those could also be quickly forgotten again.

In Monty, we started moving towards this idea with the constrained object models (Object Models) which learn the most statistically common locations and features of an object and average out irregularities or noise. However, they don’t have anything like a learning speed parameter yet (only in the sense that if the model already saw an object many times, it gets harder to make updates to it). It’s a bit like current LMs are all like the hippocampus and can learn arrangements of features and objects instantaneously. As we introduce more hierarchy into Monty, we will likely need to change this to ensure stability.

In terms of consistency and stability during inference, we have voting, where columns share their current hypothesis and can thereby quickly update each other’s hypotheses and reach a consensus. Since each individual LM evolves its hypotheses based on the observed sequence of observations, the votes themselves also contain information from past time steps and should stabilize when presented with a consistent input.

Regarding self-maintenance, structural coupling, and emergent directionality, I would have to read more about the specific definition of these concepts in the literature to be able to relate them well to Monty.

Let me know if this helps or if I can go into more detail on any of those topics specifically.

Best wishes,

Viviane

3

Thanks a lot for the thoughtful reply — this is very helpful, especially the idea of different learning speeds across modules as a way to stabilize representations over time.

Just to clarify where my question is coming from: my perspective is more rooted in systems biology and complex systems, where the focus is less on representation itself and more on the persistence of organization over time.

In that sense, I think the difference I’m trying to point at is the following:

  • The mechanisms you describe (learning rates, voting) seem to explain stability of representations and inference

  • While my question is more about continuity of the system as an organized entity

This is where I was implicitly drawing from concepts like those introduced by Humberto Maturana and Francisco Varela:

  • Self-maintenance (autopoiesis): the system actively preserves its own organization

  • Structural coupling: the system and environment co-evolve through continuous interaction

  • Emergent directionality: behavior exhibits orientation without requiring an explicit external objective

From that perspective, the question becomes more about whether there is something like a preserved organizational dynamics that defines the identity of the system across time.

I’m not sure whether this kind of perspective is currently within scope of the Thousand Brains framework, but I feel it could become relevant when thinking about larger-scale, hierarchical systems over longer time spans.

Thanks a lot!