Five years ago, i read Jeff’s papers about cortical columns and grid cells. Now it’s interesting to see the monty project. But monty uses a Cartesian reference frame, diverted from the original SDR and related simulations. I think SDR is essential to faithfully implement Human-like intelligence and wonder anybody working on this direction?
1 Like
as far as I remember, initially, TBP didn’t need SDRs but they seem to be back 
Code search results · GitHub. One can hear about them occasionally in the meeting videos on youtube
woo, the code was committed last month. they ARE back. Checking it! Thanks for the reply 
1 Like
welcome “back” too!, @Tech_LingQi
Hi @Tech_LingQi
we are currently not actively working on incorporating SDRs into Monty (but are not opposed to doing this in the future). We did explore using SDRs, HTM, and grid cells in the very beginning of Monty but hit some roadblocks. Here are two videos summarizing that project:
Since we didn’t have a use case yet where we couldn’t achieve the same functionality without SDRs it hasn’t been a priority on our research agenda. If you are interested in our general thinking around incorporating more neural elements into Monty, here is a nice summary from Niels: https://youtu.be/lewMhwzNtEo?si=BZiCMZcIDAACRnNn
One other SDR project we did (the one you saw in the code) was about using SDRs for the object IDs that learning modules output and encoding object similarity in those. Here is a video about this project: https://www.youtube.com/watch?v=qHdl8tlzY3k and documentation: Evidence as a Similarity Measure However, this is not part of our current default Monty configuration. If you have any questions around this work, @rmounir would be best to ask.
I want to highlight that just because we are currently not working on this ourselves doesn’t mean that we would want to discourage others from working on it. If you are interested in implementing an SDR-based learning module and plugging it into Monty, we’d be excited to hear about your progress and results.
Best wishes,
Viviane
2 Likes
Thanks for your reply.
i see your point: the current Cartesian representation is ok for your current goal, so SDR is not your current priority. I just think SDR and related simulations are more faithful to what’s happening in neocortex and it would be interesting to see them in action in much more realistic experiemnts, e.g., multiple 3d object recognition in Habitat, as you are doing using Cartesian representaions.
i am interested but i am not sure how much time i can spend on this topic. i am working on robotics in Isaac Lab of Nvidia. It would be fascinating to see, one day, a robot contolled by the same mechanism used by our neocortex.
Anyway, thank you for what are doing in Monty and have done in SDR in earlier days, they are interesting and inspiring. Had i built anything based on Monty or SDR, i’d share them here. This is a really nice place.
4 Likes