I would like to know some research areas in HTM.
Hi @priyojit ! Welcome to the forum!
There’s a lot of research area you can look into. Personally I work on the piratical side of things. I build high-performance HTM implementations (general optimizations, HTM on GPU, etc…) and solving Reinforcement Learning problems using HTM.
I started the Etaler for my bachelor’s degree project as a research and production orientated alternative to NuPIC and HTM.core. And designed the framework around HPC concepts. Resulting in an very fast HTM framework. You can find more details in the post announcing the first release. Even though totally beating NuPIC at performance, there’s a lot of possible optimizations I haven’t try either due to time constants or lacking equipment. I’ll release and share a project report (I’m required to do so as our university’s policy) around December detailing possible optimization strategies and other information. Please PM me if you are interested in the subject; I’m not under any NDA. You can also have a look at Etaler’s issue page to see some known optimiaztions under development.
Another part I do is putting HTM into Reinforcement Learning problems, sometimes called SMI in the community. This have been done a few times before; see here for an example. But the field is fairly young and have vast potential. The community have the 2d-object-modeling going on slow and steadily. And I’m working on applying HTM in standard RL environments (Using my framework described above).
Other forum member have their own fields. Anyone?
@thanh-binh.to may you share your project?
We are working on applying HTM to today’s machine learning systems, and looking into other cortical models. Keep up with all our research by subscribing to our YouTube channel. We publish new research videos every week.
Here is a list of our live-streamed videos, which contain a lot of research meetings.
Thanks for the info
Thanks but I have one doubt about the relationship between SDR Classifier and Neural Network.
If SDR classifier is NN, then why NN is used??
The brain theory ends in HTM with a set of cell activations that represent something. In the brain, it means something, but to an outsider, it means nothing. So we do the standard machine learning classification on it to make it mean something in the context of the input space. If you are confused about why these representations are meaningless outside the system, read Agency, Identity and Knowledge Transfer.