http://www.youtube.com/watch?v=4y43qwS8fl4

This is easily the most astounding presentation on artificial intelligence that I’ve ever seen or read before. Ten minutes into it I was already starting to rethink my entire approach to the field, and it just kept getting more and more interesting from there.

Jeff Hawkins’ talk starts out by describing what our brains really do and how they do it, and goes on to describe the mathematical models arising from those theories, the real world high value applications they have already implemented using this knowledge, and where they are headed in the future.

ABSTRACT

The neocortex works on principles that are fundamentally different than traditional computers. In this talk I will describe recent advances in understanding the neocortex and how we are applying them to model millions of high velocity data streams.

The talk will start with a description of sparse distributed representations, which are the fundamental units of information in brains. I will then discuss how these representations are learned and how the brain processes them to build predictive models from sensory data. Numenta has built a product called Grok that emulates these capabilities of the neocortex. Grok is being used to understand high velocity machine generated data in many different domains. I will give a brief introduction to Grok and speculate on the future of machine intelligence.

Andrew Smith - Mar 12, 2013:

This is easily the most astounding presentation on artificial intelligence that I’ve ever seen or read before. Ten minutes into it I was already starting to rethink my entire approach to the field, and it just kept getting more and more interesting from there.

Thanks for the link to the informative video.

It’s definitely exciting material, though (I’m trying not to brag here…) I was already familiar with most of it, especially since I’d read Hawkins’ book “On Intelligence”. Hawkins’ book was definitely an inspiration to me, especially the insight that the cortex is an organ of prediction. I believe he’s roughly on the right track, even now, though I’m harboring some insights of my own (that I’m trying to publish) that he and others don’t seem to have thought about.

The two main new things I learned from the video were: (1) there exist motor neurons in every sensory area of the brain; (2) I didn’t know Numenta was involved with Grok, and I hadn’t heard of Grok before; (3) It is believed likely that a planet exists only 13 light years away from earth. Somehow I missed that astronomical bit of news from February of this year. Here’s one online article about it…

http://www.huffingtonpost.co.uk/2013/02/07/earth-like-planet-13-light-years_n_2636992.html

But back on topic… I have one minor objection to a comment Hawkins made in that video: that our neurons learn only in online/real-time situations. It is commonly believed that one of the most important functions of sleep is to provide down time (“batch mode”) so that the brain can organize and consolidate what it has learned during the waking hours of the preceding day…

The weight of the evidence, however, seems to be coming down on the answer, yes, sleep is necessary to consolidate memories — although maybe for only some types of memory. Most of the research favoring sleep’s importance in consolidation has used procedural / skill memory — sequences of actions.
http://www.memory-key.com/improving/lifestyle/activity/sleep

Other than the few things I mentioned, I couldn’t find fault with anything he said. I agree with all his general future predictions, too. I recommend Jeff Hawkins material for anyone really serious about producing AGI.

By the way, the title of that video on YouTube, in case it disappears in the future, is:
“Building Brains to Understand the World’s Data”.

... Hence the title of my post.

I agree with the points that you are making about sleep, downtime, and batch processing.

The presenter was at pains to point out that in his talk, he had only addressed three of the six aspects of intelligence which they were dealing with at Numenta. Furthermore, he acknowledged in his response to one of the questions at the end that there were other elements that would be necessary for AGI which were outside the scope of their project altogether (e.g. goal management).

Edit: I’m also wondering if this revelation that was announced recently has any bearing on what they are doing at Numenta.

http://www.chatbots.org/ai_zone/viewthread/1190

Interesting talk indeed!

We had been trying to determine the best way to do something like this ourselves, being able to represent arbitary data, without any external definitions within Caesar.

Some idea’s we had were workable, one in particular was “similar” (which I use loosely) to Hawkins’s method.

However, after watching that video, it helped to fill in a few gaps in our idea’s, probably because none of us here are neurologists per se.  Amazingly Hawkins’ model here fits almost perfectly into ours and we should be able to move forward and save any further headscratching.

There were a few things I didn’t agree with also, the batch mode being one of them, and some idea’s we have had since will address some of the shortcomings of thier implementation.  Specifically, the need for the user to define the data and provide “encoders”. 

Perhaps they didnt look at these themselves yet as they weren’t affecting the final goal for them, but for us, we require a completely none-supervised system, that doesn’t rely on any external definitions and moving forward from Hawkin’s presentation we may be able to address that.

For us now, we are developing our own “HTM” using idea’s we have had from this presentation, coupled with “improvements” to address the shortcoming for our application.

If anyone’s interested I’ll post periodic updates on how we are doing and what we have done.