Basic Question on thinking feedback loop and the HTM

I have been looking at Numenta work off and on for years but just basic amateur level. One thing, I like the building of the model and the thousand brains. But it seem that the AI built will rival… I guess image recognition models that are out there now possibly. Will this approach lead to more general artificial intelligence, the kind that thinks and constantly operating.

For example, even with ChatGPT, you give a “A prompt” and it responds with data. Humans live 80, 90, 100 years, always responding to its environment. Do you have to facto that into systems like Numenta where data is always changing the state of the brain. Is there more research on that aspect? information coming from outside of the brain? And do you replicate that and how?

1 Like

I can’t speak for Numenta or what they’re currently working on in their lab.

I see HTM with TBT as an unsupervised training - spatial/temporal pattern completion system.

On its own, a single HTM/TBT network has limited practical applications.

Combining multiple HTM/TBT sub-networks could replicate the functions of a transformer network, offering advantages like sparsity and distributed information. This could allow for parallel processing, online training, and the addition of temporal elements. Various HTM/TBT sub-networks could be arranged in the architecture envisioned for the global workspace. These sub-networks might work together to find a low-energy global solution, serving a similar purpose to transformer attention heads.

The optimal configuration of feed-forward & feedback (context/location) connections is a huge unknown. I can see the context path as an efficient interrogation path.

Unlike transformers, which require extensive batch training, HTM/TBT might support incremental and unsupervised training.

In the forum and Numenta’s published work, there seems to be little focus on the ‘H’ in HTM. However, I believe that exploring this aspect is the way forward.

1 Like

I didn’t quite get your questions but for this particular line - I would say YES.

The algorithms in HTM (SP and TM) constantly update their models in real-time as it ingests inputs. Of course these updates can be turned off if not desired.

The “state of the brain” can be thought of the configuraiton of the mini columns in an HTM model, and they get updated in real-time as mentioned above.