# 2026/08 - Rome Focus Week Final Presentations **URL:** \ **Category:** Video Discussions\ **Tags:** brainstorming-video\ **Created:** [September 29, 2026, 3:04pm UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215 "2026-09-29T15:04:49Z")\ **Posts on this page:** 12\ **Page:** 1
**Author:** ![TBP-Announcements](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/tbp-announcements/32/1065_2.png) [@TBP-Announcements](https://forum.thousandbrains.org/u/TBP-Announcements)\ **Post date:** [September 29, 2026, 3:04pm UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/1 "2026-09-29T15:04:49Z")
The team presents the results from the Rome Focus Week Retreat. Team Mighty Mouse demos their ARC-AGI-3 integration, Team Fate Attenzione demos their new attention system prototype in Monty, and Team Janus demos a new ability to merge or split objects in Monty. # Main Video [![](https://canada1.discourse-cdn.com/flex034/uploads/thousandbrains/original/2X/e/e61e957d8ac3d20f5292bd6ff681df3a6d5f9dc7.jpeg "2026/08 - Rome Focus Week Final Presentations") ](https://www.youtube.com/watch?v=QsTJY-83EUg) [0:00](https://www.youtube.com/watch?v=QsTJY-83EUg) Introductions [0:10](https://www.youtube.com/watch?v=QsTJY-83EUg&t=10s) Rome Focus Week Recap [2:48](https://www.youtube.com/watch?v=QsTJY-83EUg&t=168s&pp=0gcJCWMAwfN6Pr3D) Team Mighty Mouse: ARC-AGI-3 (Dale, Jeremy, and Hojae) [4:45](https://www.youtube.com/watch?v=QsTJY-83EUg&t=285s) ARC-AGI-3 Game Demo [7:04](https://www.youtube.com/watch?v=QsTJY-83EUg&t=424s) Game Engine API for Custom Tasks [8:40](https://www.youtube.com/watch?v=QsTJY-83EUg&t=520s) ARC-AGI Simulator [10:30](https://www.youtube.com/watch?v=QsTJY-83EUg&t=630s) Sensor Module to Detect Changes [15:20](https://www.youtube.com/watch?v=QsTJY-83EUg&t=920s) Learning Compositional Game Maps (and Sprites) [21:50](https://www.youtube.com/watch?v=QsTJY-83EUg&t=1310s) Mighty Mouse Conclusions [23:47](https://www.youtube.com/watch?v=QsTJY-83EUg&t=1427s) Mighty Mouse Q&A [31:04](https://www.youtube.com/watch?v=QsTJY-83EUg&t=1864s) Team Fate Attenzione: Building an Attention System in Monty (Scott, Viviane, Tristan) [44:21](https://www.youtube.com/watch?v=QsTJY-83EUg&t=2661s) Model-Free Attention Prototype [46:10](https://www.youtube.com/watch?v=QsTJY-83EUg&t=2770s) Model-Based Attention (Including Attention Inhibition) [48:29](https://www.youtube.com/watch?v=QsTJY-83EUg&t=2909s) Filtering Percepts [53:57](https://www.youtube.com/watch?v=QsTJY-83EUg&t=3237s) Adding a Background & Removing “Off Object” Filtering [58:39](https://www.youtube.com/watch?v=QsTJY-83EUg&t=3519s) Adding Soft Attention & No-Match Detection [1:01:05](https://www.youtube.com/watch?v=QsTJY-83EUg&t=3665s) Salience Change Detection & Explored Tracking Moving Objects [1:07:46](https://www.youtube.com/watch?v=QsTJY-83EUg&t=4066s) Dynamic Voxel Sizes [1:09:06](https://www.youtube.com/watch?v=QsTJY-83EUg&t=4146s) Fate Attenzione Summary & Q&A [1:14:50](https://www.youtube.com/watch?v=QsTJY-83EUg&t=4490s) Team Janus: Object Merging & Splitting (Ramy, Benji & Niels) [1:21:02](https://www.youtube.com/watch?v=QsTJY-83EUg&t=4862s) A New Web GUI Visualization [1:21:49](https://www.youtube.com/watch?v=QsTJY-83EUg&t=4909s&pp=0gcJCWMAwfN6Pr3D) Web GUI Demo [1:23:13](https://www.youtube.com/watch?v=QsTJY-83EUg&t=4993s) Object Merging [1:31:57](https://www.youtube.com/watch?v=QsTJY-83EUg&t=5517s) Object Merging Experiments (Using Model Ablations) [1:37:51](https://www.youtube.com/watch?v=QsTJY-83EUg&t=5871s) Object Merging: Neuroscience Discussion [1:42:20](https://www.youtube.com/watch?v=QsTJY-83EUg&t=6140s) Object Splitting [1:44:27](https://www.youtube.com/watch?v=QsTJY-83EUg&t=6267s) Splitting Demo & Results [1:50:09](https://www.youtube.com/watch?v=QsTJY-83EUg&t=6609s&pp=0gcJCWMAwfN6Pr3D) Object Splitting: Question about Bursting [1:54:10](https://www.youtube.com/watch?v=QsTJY-83EUg&t=6850s) Object Splitting Implementation Details [1:55:32](https://www.youtube.com/watch?v=QsTJY-83EUg&t=6932s) Future Work & Open Problems [1:57:31](https://www.youtube.com/watch?v=QsTJY-83EUg&t=7051s) Judges Interactive Demo: Monty in Browser [2:05:12](https://www.youtube.com/watch?v=QsTJY-83EUg&t=7512s) Closing & Judges Deliberation ---
**Author:** ![AgentRev](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/agentrev/32/651_2.png) [@AgentRev](https://forum.thousandbrains.org/u/AgentRev)\ **Post date:** [October 5, 2026, 6:43am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/2 "2026-10-05T06:43:40Z")
Incredibly interesting projects from all the teams yet again! 😃 About the ARC 3 test, even if it wasn’t meant to yield an immediate result, forcing to think differently is really the main benefit, as Hawkins mentioned in a past meeting. It’s also good preparation for the next ARCs. When you look at Astra’s replays, it’s poking things around and assigning semantic tags to the sprites based on their behavior and reactions, allowing the LLM to leverage all the pre-trained behavioral knowledge it has about those tags. It hints at abstract higher cognition being an initial requirement to properly recognize behaviors and affordances. (This doesn’t require semantic tags, but at the very least dimensionally-reduced prior knowledge of the involved concepts.) The attention system is my personal favorite. I was initially a bit skeptical of the approach as the presentation started, but my few concerns were all addressed by the end of it. Very neat. I guess a more complex case to visit in the far future would be context-based attention, e.g. if I see an insect flying indoors from the corner of my eye, it will immediately grab my attention, but if I see one outdoors, I will likely ignore it. Well, unless it’s a wasp… The web-based GUI is plain awesome. This is exactly the kind of stuff that can attract more people to the project. Great work. Merging and splitting are well thought-out all around; quite clever to use sensorimotor exploration directly. Hawkins never stops throwing his insightful edge cases into the fray 😆. I guess some aspects of those edge cases would have to be addressed by abstract higher cognition, e.g. a unique object successfully keeping a consistent UUID over its lifetime even as its physical properties evolve (like losing a chip / being worn down). ---
**Author:** ![greg](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/greg/32/1623_2.png) [@greg](https://forum.thousandbrains.org/u/greg)\ **Post date:** [October 7, 2026, 1:28am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/3 "2026-10-07T01:28:16Z")
Great presentation! What if instead of this idea of merging and splitting models you have a neural network responsible for pointing to different existing models? Why is a new model being created in the first place to merge or split? Why not have models themselves be defined as a thing. I.e. a handle is patch of neurons that kinda look like a handle, and there’s a hierarchy built out of it… and to add a fork or a spoon is just a neighboring network that connects to the handle? ---
**Author:** ![nleadholm](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/nleadholm/32/48_2.png) [@nleadholm](https://forum.thousandbrains.org/u/nleadholm)\ **Post date:** [October 9, 2026, 5:04pm UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/4 "2026-10-09T17:04:07Z")
> [@greg](#): > > neural network responsible for pointing to different existing models Thanks for your question Greg. Re. your first question - it’s worth highlighting that in unsupervised learning, one of the key items is that a learning module (which is meant to represent a cortical column) tries to infer the object and match it to its existing models. It is only if it is unable to match it that further unsupervised learning is necessary, either the creation of a new model, or as we show here, splitting the model. Monty does not use “neural networks” (in the deep learning sense), as it is built on a different view of how biological intelligence works. Does that answer your question? I’m not sure I followed your other comments, but perhaps if you can elaborate then I can respond further. ---
**Author:** ![greg](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/greg/32/1623_2.png) [@greg](https://forum.thousandbrains.org/u/greg)\ **Post date:** [October 9, 2026, 6:57pm UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/5 "2026-10-09T18:57:42Z")
> [@nleadholm](#): > > Monty does not use “neural networks” (in the deep learning sense), as it is built on a different view of how biological intelligence works I was under the impression that Monty was based on how biological intelligence works, as you say, so why do you say it doesn’t use neural networks, as biological intelligence works using neural networks? I think my message might have been misunderstood as it was probably poorly written due to general ignorance of this field (it’s been a while since I’ve dived back into it). Let me try again. The idea in my head is that neural networks would form automatically based on sensory data to represent stable concepts (“models”). These stable concepts could link to each other. Other neural networks would be made out of them, and whenever a new stable concept came, it would be built out of the components of similar previous models. This wouldn’t require merging or splitting, but rather pointing. So say I have a mug without a handle: ```auto [Mug without a handle] -> [liquid] + [container] ``` Then when I encounter a mug with a handle it’d be like this: ```auto [Mug with a handle] -> [handle] + [mug without a handle] ``` The sensory data somehow would get sorted automatically to prexisting concepts or create just the small bit of new concept and link it to old concepts. Does that make sense? This is a different idea from the idea that there might be multiple identical models that need to be “merged” or “split”. As Jeff Hawkins said, it’s unlikely the brain uses “merge” and “split” operations. ---
**Author:** ![greg](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/greg/32/1623_2.png) [@greg](https://forum.thousandbrains.org/u/greg)\ **Post date:** [October 9, 2026, 8:46pm UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/6 "2026-10-09T20:46:47Z")
Taking this further with the example from the video: ```auto [broken mug with handle] -> [broken] + [mug with a handle] ``` Again, no need to merge or split. Just point. ---
**Author:** ![rmounir](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/rmounir/32/268_2.png) [@rmounir](https://forum.thousandbrains.org/u/rmounir)\ **Post date:** [October 12, 2026, 12:53am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/7 "2026-10-12T00:53:42Z")
Thanks @greg. I have a few clarifying questions. - How would you represent these pointer operations in the brain? I think part of the confusion about the term “neural networks” is that it is missing the details of how this would be represented in the brain, and whether you are referring to a biological neural network or an artificial neural network for learning these pointer operations. - Are you thinking a compositional model where `Mug with a handle` is in a higher-region column with lower-region child objects, e.g., `handle` and ` mug without handle`, send their object ids to the higher-region column? - Are you thinking of a union of SDRs being active in the same layer, where each SDR represents some concept/object, e.g., `handle` and `mug without a handle`? In other words, you combine these concepts in additive manner to represent the whole object without the use of hierarchy? - How do these stable concepts get created and updated over time? This project of merging and splitting is addressing the problem of how to discover and continuously update these stable concepts/objects from online sensorimotor data. - Seeing the mug without a handle first, then seeing the mug with a handle is a simplification of the problem that requires supervision. If you instead see the `mug with a handle` first, wouldn’t you need to split it into a `mug` and a `handle` the next time you see any of these objects as part of different models, e.g., same handle on a different object? - But even for this simple case of seeing `mug without a handle` first, then adding a `handle` to it as a separate object when you see a `mug with a handle`, why did you choose to add the handle as a separate component (and point to it) rather than simply extending the `mug without handle` object to include the `handle`? I think the challenge is in figuring out what makes a “stable object/concept”, this project proposes that it is an online process that gets refined over time as you see other objects that share common child objects. The proposal here is that this could happen through continuous merging and splitting of objects. ---
**Author:** ![greg](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/greg/32/1623_2.png) [@greg](https://forum.thousandbrains.org/u/greg)\ **Post date:** [October 12, 2026, 3:26am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/8 "2026-10-12T03:26:15Z")
My understanding was/is that Numenta is interested in starting with the brain as its primary inspiration and model for machine learning. So when I say “neural networks”, I am referring to neural networks in the brain, and whatever attempts to mimic them Numenta is doing. According to neuroscience, the brain forms hierarchies out of basic features and builds from basics to complex concepts. So the mug would be no different. Both the mug and the handle would be built out of basic shape networks, all pointing to each other. These would be tied in to behavioral networks connected to limbs. The associations between different behaviors (like grabbing, pulling, etc.) would all dynamically be built together somehow. A handle and a doorknob would get links together because the fingers grab both. The word “handle” is a sound that links with grabbing and the object handle too. The arrows in my drawings are just crude representations of the firing of synapses between these networks. I haven’t studied any of the work that Numenta has done since looking at HTMs 13 years ago. So I don’t know what SDRs are. But it seems to me that it’s unlikely that the brain produces new models all the time that are in need of merging/splitting. Does a new cluster of neurons suddenly pop into existence each time you look at something? Do clusters of neurons get eliminated in the brain when it realizes they’re similar? I don’t think the brain works like that. Somehow the brain encodes concepts out of small features in hierarchies, and these concepts can be objects, motions, sounds, etc., and they all point and link to each other. How it does this, I have no idea. I figured that’s what you guys are trying to figure out. Of course, as I am a mostly ignorant person when it comes to this subject, we might be talking past each other. My only reason for speaking is in case something I’ve said might be useful, as you guys have been knee deep in the trenches and I have the “Beginner’s Mind” of knowing little. ---
**Author:** ![rmounir](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/rmounir/32/268_2.png) [@rmounir](https://forum.thousandbrains.org/u/rmounir)\ **Post date:** [October 12, 2026, 6:29am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/9 "2026-10-12T06:29:00Z")
> My understanding was/is that Numenta is interested in starting with the brain as its primary inspiration and model for machine learning. Yes, you’re right. Both, Numenta and the Thousand Brains Project, are interested in building AI systems based on neuroscience principles. Thanks for clarifying. > the brain forms hierarchies out of basic features and builds from basics to complex concepts. In the theory, we describe the hierarchical processing in the neocortex differently. Here the hierarchy represents the compositional associations between full objects, as opposed to the common view of detecting simple-to-complex features at the different levels of the hierarchy. So, a `mug with a handle` that splits into `mug without handle` and `handle`, could still form a compositional `mug with a handle` object by associating the recently split child objects to the parent object model. If we have different types of mugs that use the same handle, the compositional association would allow for learning different types of mugs without re-learning the same handle object in each of them. The below image is from [a recent TBP paper](https://arxiv.org/pdf/2507.05888) ![Screenshot 2026-10-12 at 8.04.07 AM](https://canada1.discourse-cdn.com/flex034/uploads/thousandbrains/original/2X/5/5037d247313a08aa32a95ef7eec7eb12154652a1.jpeg) > I haven’t studied any of the work that Numenta has done since looking at HTMs 13 years ago. So I don’t know what SDRs are. Sparse Distributed Representations (SDRs) is the data structure used in HTM. It’s basically just a large binary vector with a sparsity constraint, i.e., only a few bits are active at any time. > Does a new cluster of neurons suddenly pop into existence each time you look at something? Do clusters of neurons get eliminated in the brain when it realizes they’re similar? I don’t think the brain works like that. These are interesting questions. The way I see it is that when we merge or split we do not re-learn the features at location associations, i.e., L4 \<—\> L6 synapses. Those features at location associations are the actual detailed graphs. Instead, what is re-learned is the association of the graph with an object ID in L2/3. It is more like re-labeling the already learned objects, or parts of them. Note that this was a one week hackathon research prototype for an idea that we have not sufficiently discussed in our research meetings beforehand, so there may be some details here that could turn out to be incorrect. ---
**Author:** ![nleadholm](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/nleadholm/32/48_2.png) [@nleadholm](https://forum.thousandbrains.org/u/nleadholm)\ **Post date:** [October 12, 2026, 9:55am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/10 "2026-10-12T09:55:29Z")
Thanks for the comment Greg. To add to Ramy’s input, I thought I would clarify further re. neurons suddenly “popping into existence” - it’s worth making a distinction between how Monty currently works, and how the Thousand Brains Theory predicts cortical columns would work. In the brain, we have grid-cell-like modules that have a fixed capacity - the same set of neurons that need to be re-used to represent everything. When a new reference frame is created, new neurons are not created - it is the unique phases of the grid cells that uniquely separates out this reference frame from others. [This video](https://www.youtube.com/watch?v=mP7neeymcUY&list=PL3yXMgtrZmDqhsFQzwUC9V8MeeVOQ7eZ9) might be helpful if you’re new to grid cells. In the case of Monty, our current emphasis is to keep things simple and interpretable, so we just instantiate a new Cartesian space in computer memory. In the long term, I think we are likely to adopt a distributed, fixed capacity code similar to grid cell modules. However, we may still create neurons “on the fly” as we do when expanding the hypothesis space during inference. ---
**Author:** ![greg](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/greg/32/1623_2.png) [@greg](https://forum.thousandbrains.org/u/greg)\ **Post date:** [October 12, 2026, 5:36pm UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/11 "2026-10-12T17:36:32Z")
Thank you @nleadholm! I appreciate your reply and @rmounir’s and need to make more time to dive deeper into this stuff again. It’s so fascinating to me but there are just so many hours in the day that I can give time to all my various interests. > [@nleadholm](#): > > [This video](https://www.youtube.com/watch?v=mP7neeymcUY&list=PL3yXMgtrZmDqhsFQzwUC9V8MeeVOQ7eZ9) might be helpful if you’re new to grid cells. Ah this brings me back. What a great video, thank you! It’s so sad that Matt’s no longer with us, he did such an amazing job putting out educational material. I am forever grateful. 🙏 ---
**Author:** ![nleadholm](https://yyz2.discourse-cdn.com/flex034/user_avatar/forum.thousandbrains.org/nleadholm/32/48_2.png) [@nleadholm](https://forum.thousandbrains.org/u/nleadholm)\ **Post date:** [October 13, 2026, 11:14am UTC](https://forum.thousandbrains.org/t/2026-08-rome-focus-week-final-presentations/1215/12 "2026-10-13T11:14:53Z")
No worries at all, really glad it was helpful, and please don’t hesitate to ask more questions as you dive deeper. And yes I totally agree. I never had the privilege of meeting Matt in person, but like many people on here, his videos were fundamental to my first exposure to HTM. A wonderful person.