HTM Hackers' Hangout - Mar 2, 2018

live-event
community

#1

9AM Pacific time! Suggest topics in a comment below!

HTM Hackers’ Hangout is a live monthly Google Hangout held for our online community. Anyone is free to join in the discussion either by connecting directly to the hangout or commenting on the YouTube video during the live stream.

If you have something specific you’d like to discuss, or if you just want to learn more about the HTM Community, please join HTM Forum at https://discourse.numenta.org. We have active discussions about HTM theory, research, implementations, and applications.

More info on all these topics at http://numenta.org.

I plan on talking about:

  • MIT AGI Course
  • BAMI format and possible future work
  • Getting rid of Slack
  • Unity and HTM (hopefully @lscheinkman can join)
  • research update (hopefully @mrcslws can join) (all our researchers will be at COSYNE)

#2

Topic ideas:

  • Thoughts on innateness vs learned capabilities. Do you believe that the human brain comes hard-wired with some capabilities “out-of-the-box,” per se? e.g. basic object recognition, such as faces or edge detection. If so, how does HTM plan to address innateness that humans may have as the product of evolution?
  • I’ve heard you talk about distal dendrites. Is that the same as apical dendrites in pyramidal neurons?
  • How does HTM model neurons? e.g. Are they summing machines similar to many classic neural networks, do they have a bio-realistic time component, does the output have bio-realistic polarization states (the up, down, up rollercoaster shape), are there different dendrite groups, are there different types of neurons in the HTM model?
  • How do you measure that changes to HTM are useful? e.g. if one of your researchers read a neuroscience paper on a new way to do synaptogenesis, and it was implemented in the HTM code, how would you test that those changes were useful?

Thanks!


#3

I can answer all those, thanks!


#4

How a field of HTM neurons recognizes a complicated pattern that is presented to the entire ensemble at the same time.

Try one novel, one a partial match, and one learned pattern and then the first novel pattern after the initial presentation and show how the network responds to each presentation.

By observing the network how do I know if it seeing a match, partial match, or completely novel thing?


#5

Not sure I can express that well without some coding, and seeing how I only have a couple hours I don’t think I can do it. Will try to talk about it.


#6

Great, see you on YouTube!


#7

I will post a link to join the hangout live in about 30-40 minutes. I won’t post it on twitter this time because thar be trolls on twitter. I forgot to post the link to watch, so here it is:


#8

Join here: https://hangouts.google.com/hangouts/_/wjlbfiuzfva3lmw3ynw33amnjme

Starting in about 10 minutes


#9

Rats - hangouts seems to be blocked by the company firewall policy at work.
I will just have to watch the youtube thing later.
Grrrrrrr.


#10

Innateness: Sounds like you lean towards mostly learned and not as much innateness. Gary Marcus’s book Birth of the Mind presents interesting examples of babies that can do things without learning; that the baby mind is not a blank slate that learns everything but that has things built-in during growth in the womb.


#11

Dendrites: Excellent explanation. In short, you have dendrites split into apical, basal-proximal, and basal-distal. Focusing on proximal and distal.


#12

Q: How does HTM model neurons? A: HTM is a Hebbian learning system. No bio-realistic time component. Take input one after the other without any timing. There is a video out there about timing and oscillations and that is discussed. All neurons are parametal neurons - grouped into layers. Might have grid cell modules in the future. Have inhibital neurons (not instances of, but model the effect of inibitial neurons). Inhibition is done in a more simplified way.


#13

Q: Tests and benchmarks. A: Don’t change the core of HTM theory often at all. e.g. distral [basal]-dendrites and their predictive state. Do all of the messing around in research repositories (RRs). Nupic is stable. In RRs would fully vet out ideas and that logically it makes sense. Then, have benchmarks and tests they run. Takes a lot of vetting to get into official code base of Nupic.


#14

Sorry for the confusion, the comments I just posted were not additional questions, but merely summaries of your answers. Wasn’t trying to be that guy that feels like everything he asks from his unlimited well of questions demands everyone’s attention. :smile: Thanks!


#15

Peace out


#16

@mike2.0 thanks for your summaries of my answers!