Jeff, Vivian and the Thousand Brains Team come together to ask questions around Monty, learning modules, human wiring, the hippocampus and more.
Watching the first part of this video, I realized that I only recognized a few of the folks on the team. How about setting up “bio” pages for each of the team members, coupled with an index page containing names, precis, and thumbnail photos? A link to the index page could be added to each video’s cover notes.
1 Like
It’s a good idea. This is the bio page - Our Team | Thousand Brains Project
1 Like
I just checked out the page. It’s nicely done and pretty much what I was asking for. However, I have a small nit to pick: The popups that appear when I click the “More info” buttons have extremely small type, making them needlessly hard to read. Of course, I can zoom the screen, but maybe they should start out at a larger size…
Although I understand LM are meant to each develop some model of the object/world (consistent with the type of sensory modules connected to it), I think discussions focussing on the perspective of a single learning module in isolation sometimes misses important features of the system. Functioning brains always have large number of LMs.
Individual sensors read local features but poses and locations are stable and reliable because they arise from the voting of “many” .
Discussing allocentric / bodycentric and coordinate-transform as if all the information must be gathered and processed from within a single learning module suggest something hugely resource-intensive for the LM; possibly not the most productive avenue to use for your project in the long run.
Hi Jacques,
We completely agree that understanding the full functionality of the neocortex is a critical goal. However, through years of research in these complex areas, we’ve found that focusing initially on simpler structures is the most effective approach. When considering the tens or hundreds of factors introduced by hierarchy and cross-connectivity, the search space for a correct solution becomes overwhelmingly vast. By studying these systems in isolation, identifying and validating the correct solution becomes significantly more manageable. With a solid foundation in place, incorporating additional factors from the connected neocortex then becomes a much more streamlined process, dramatically narrowing the search space.
We also believe that stability can be achieved with a single learning module, provided it is paired with multiple movements of its associated sensor. A thought experiment we quite often do is imagining looking at the world through a straw, and how that tells us that a single cortical column must already be very powerful and implement all the critical mechanisms.
While we agree that performance will be an important consideration in the future, our current priority is to make significant progress on key roadmap items. That said, we are mindful of potential performance impacts and regularly monitor our benchmarks.
2 Likes