How learning in monty works (algorithmically)?

Monty Code
1

Good day,

Firstly, my apologies if this is already a redundant question.

I would like/need to understand how learning works algorithmically for monty. How?

I’ve watched a couple of videos already and tried to grep in this forum, but unfortunately I couldn’t find any document/article relating to this, nor I understood the algorithm. I do plan to dive into code, but it’s great to have a good background first. I’m referring to the actual learning algorithm, for example as an analogy, in ANNs it’s through backpropagation (gradient descent). In HTM spatial pooler it’s through hebbian learning, permanence values of synapses are updated. For monty, I know it’s using reference frames and specifically object poses and features - how do these achieve learning?

1 Like
2

Hey there @pepedocs

It’s been a minute since I studied Monty’s algorithm up close, so my understanding may not be completely up-to-date, but the short answer is that it’s learning uses a mixture of graph construction and probabilistic hypothesis scoring (which is all pretty different from weighting). It encodes how features connect across poses within a reference frame, reinforcing it’s graph structure through a process of repeated prediction/verification.

I don’t want to to dive too far into this without catching myself up on how the architecture has changed in the last 6 months (don’t want to steer you wrong), but the above is a super high level description of what’s going on.

Hopefully a more recent member can step in and correct me if I’m wrong (@AgentRev in particular seems up to speed). Otherwise, I’m sure the Monty dev team will explain. Their community involvement’s always been top tier.

2 Likes
3

Thanks @HumbleTraveller for the insight. For some reason it gave me the impression (after reading your response) that it is doing some kind of Bayesian network inference. But I don’t know yet.

a process of repeated prediction/verification.

Maybe this is the hebbian-like learning part.

1 Like
4

If I’m remembering right, it’s been described as hebbian based learning to me before. So you might not be too far off base in thinking that.

Also, I managed to dig up an older thread talking about something similar. It was focused on something slightly more specific than what I think you’re asking, but even still, it might prove helpful to you.

1 Like
5

Hi @pepedocs

I think the best place to start if you want a mathematical description of Monty is this preprint: [2507.04494] Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference It formalizes Monty’s learning and inference algorithm in detail.

If you prefer a slightly higher-level description, the “How Monty Works “ section in our documentation may be a good place to start. In particular, the page on the evidence-based learning module (which is the current default).

You could call the learning “Hebbian-like” as observed locations and features are associated instantaneously. We also have object models that take into account the frequency of observing features at certain locations and hence store the most consistently observed features at locations (more details in the page on object models).

It’s not really Bayesian network inference, as we have no global normalization, but instead rely on local evidence accumulation to make it more biologically plausible.

I hope this helps!

  • Viviane
5 Likes
6

Thank you so much these are very helpful @vclay . I will definitely have a read on all of them.

1 Like