HTM Chat with Jeff



This recording went pretty long, so I broke it into a playlist. Hope you enjoy it!

I tried to hit a lot of the points brought up in Questions for Jeff?, but I’m sorry I could not get to all of them.

Does numenta do "constraint satisfaction"?
How are cellular layers 1-6 modeled in HTM?
Preliminary details about new theory work on sensory-motor inference
A different point of view on building AI system
Timing Circuits
Exploring Reinforcement Learning in HTM
Any questions for Jeff?

First, this is really fantastic, you did an amazing job. Second, chunking this was definitely the right thing to do, it makes it MUCH more approachable. Third, many thanks to Jeff for his openness and time!


Very good! I am really excited about the level of detail on sensory motor integration from the white boarding and slides – this can be easily taken to implementations of various parts of the theory. Video 1 starting around 9:11 for example is relevant to the RL app I’m working on… I had come to a slightly different design (output from sequence layer as input to motor layer) but it makes much better sense with the simple arrangement Jeff has described. Hopefully I’ll have some more experimental demos to post here on the forum as I dive into the new concepts.


Great conversation, really enjoying to watch. I’m looking forward to read upcoming paper. I want to comment on the last two videos of the series.

Firstly, I totally agree with your expressions in the beginning of 8th video, I thought the similar things, like describing life or evolution as an information accumulation process, after reading “The Selfish Gene” from Richard Dawkins.

As Jeff stated, we are tend to think about new technologies as in the way we currently know but that’s not the reality. It seems like we are also tend to interpret things in human-centric way as a matter of course. I guess, in short-term, HTM technologies will be used for enhancing the way we control and understand the world. However, in my opinion, the real beauty and importance of HTM theory is to build new intelligent forms so that they can discover and understand the universe better than us. It will be the next breakthrough after homo sapiens and end the sovereignty of biological forms on information accumulation.



Great job! It was wonderful to hear more detail on newer additions to the theory, and awesome job coordinating, producing and making this all possible for the community - much appreciated!


An excellent conversation.
I certainly hope Jeff enjoys doing these, as much as we enjoy watching them.

Please let him know it is very much appreciated. I’m just as much enamored with
this subject (Theoretical framework of brain function) as he is!

Been following Jeff’s talks & interviews for quite some time now. Since my first
encounter with the Berkeley 2012 interview: “On Intelligence with Jeff Hawkins -Conversations with History”
…and so many lectures, talks, and presentations he’s done along the way.
(TED, GoogleTalks, Berkeley-talks, Neuro-conferences, etc.)

So I share his passion, for this area of Neuroscience.

With all of the progress made up to this point, I’m fully on-board with Numenta’s mission, and am
proud (and grateful to participate) to share the vision of developing true Neo-cortical AI.

As I’m studying the published papers, and working towards making contributions to the work myself,
I think we’re still only in the earliest stages of locking down our understanding, and then translating
this knowledge to the information-processing environment. What you have now (i.e. Nupic) is foundational, and will be refined & augmented. Then, I expect , will be ready to scale, to do amazing things in AI.


On a lighter note :sunglasses:@rhyolight:

You’re certainly becoming quite the production talent in video presentations.
(a.k.a. HTM-School ) Hasn’t gone unnoticed.

I see you’re the musical enthusiast, so you I can appreciate the mixing you do with your video work …

See if this song sticks in your head like me, after having watched this presentation …

Enjoy-- :slight_smile:

( Sounds best, played at ~ 0.9 speed – really! )


A post was merged into an existing topic: Preliminary details about new theory work on sensory-motor inference


Mike that’s one of the best songs. :heart_eyes:


A post was merged into an existing topic: Preliminary details about new theory work on sensory-motor inference


And he surely does! Thanks Jeff and Matt for the time and effort. I picked up new things to chew on.


Watching HTM school and seeing the sequence memory slides, one could form the idea that a column represents a bit in a SDR, but the 2-layer model in part 3 seems to imply that a single column holds fluctuating interpretations (L4) and stable object representations (L2/3), which through feedback semantically collapse or coerce the interpretations.
Which is also the job of the higher regions? (I take it the difference is abstraction in the region- vs. structural and temporal integration in the inter-layer feedback.)


Since the output, which are stable representations of objects, are theoretically what get sent to higher regions, this would seem to imply that the higher regions (being the exact same type of circuit as the lower regions) would see those objects as now themselves the features of objects (those objects would just represent things with a higher order of complexity in the real world). Note that the word “object” can refer to abstract concepts - doesn’t have to be something you would traditionally consider an “object”.


There are two types of columns, minicolumns and macro columns a.k.a. cortical columns (CCs) The new type, the CC, is a lot bigger and contains minicolumns. I think you are right about both columns, it’s just that they’re different types. I’m not sure if the output layer is made of minicolumns.

To add onto what Paul Lamb said, higher regions can also recognize things which take up more of the visual field. There are lots of other possible functions, such as specialization (right parameters and inputs to best process e.g. speech, or processing a specific combo of senses) or higher functions (memory, awareness of emotions, etc.)


Thanks. I was just pondering this seeming overlap in functionality.
(Pardon if this is discussed in On Intelligence, it’s a while since I’ve read it) -
Regarding specialization in the Brodmann sense, and when talking about a universal cortex algorithm/structure, indeed the question comes up what constitutes e.g. sensorimotor input (efferent copies) or motor output for the higher regions and if this is recognizable in the layer structure. Even for the lower regions - what would be the meaning of the motor layer in the olfactory cortex or Wernicke’s area.