Hacker News new | past | comments | ask | show | jobs | submit login
Beating Atari Pong on a Raspberry Pi without backpropagation (ogma.ai)
63 points by phonebucket 12 days ago | hide | past | web | favorite | 5 comments





This company does very intriguing work, with some inspiration from numenta, as far as I know. The low resource usage (running on CPU, on a raspi!) is quite impressive. I wish there was more research on this approach.

This looks very interesting but whitepaper is completely offputting. It's filled with terms you would have never heard elsewhere but authors use them without much explanation. It's really not even a whitepaper, just bunch of half done slides. It's surprising that authors would so much time around writing code, taking videos etc but have no time to really write good description of their ideas. There got to be a simple and better way to explain how exactly their proposed system work.

Reminds me of some of the stuff Vicarious is doing.

https://www.vicarious.com/posts/general-game-playing-with-sc...


Mixing backpropagation and gradient descent does not inspire a lot of confidence.

What they are describing seems to be based on hierarchical temporal memories (https://en.wikipedia.org/wiki/Hierarchical_temporal_memory)


I once bet Pong on a Atari 800 XL!



Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: