Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The Grok prediction engine from Numenta announced

numenta.com

Post reply on HN
Print view View original

21–26 of 26 posts

Re: The Grok prediction engine from Numenta announced

#21

Post by ScottBurson »

sown wrote: post #9

So some ML/AI researchers I know consider Numenta to be of a somewhat "outsider artist" effort. Their opinions of the research from Numenta is that of lower regard. Now, you don't have to upvote me for this, but you don't have to downvote me, either, because I happen to know experts in ML/AI who disagree with Jeff Hawkins. Got that? Downvoting me doesn't change these other people's opinion. I'm just relating to you w…

What you have to factor in here is that people who consider themselves experts will often take a dim view of something new they don't understand, particularly if the work was done outside the system they're part of.

(A quote comes to mind: "The bomb will never go off, and I speak as an expert in explosives." -- Admiral William Leahy to Harry Truman, concerning the atomic bomb.)

Numenta's technology is definitely novel. How well it will work remains to be seen, but I think it's a very interesting experiment.

Re: The Grok prediction engine from Numenta announced

#22

Post by Element_ »

reader5000 wrote: post #13

I'll remain skeptical until it is determined how their "cortical algorithms" benchmark compared to, say, Naive Bayes. Good marketing though.

Last year Dr. Davide Maltoni published a paper (Pattern Recognition by Hierarchical Temporal Memory)[1] comparing the algorithms laid out in Dileep Georges original PHD thesis. The results were pretty impressive. There have been many improvements to the algorithm since then too.

http://bias.csr.unibo.it/maltoni/HTM_TR_v1.0.pdf

Re: The Grok prediction engine from Numenta announced

#23

Post by X4 »

msellout wrote: post #10
wiradikusuma wrote: post #7

How is it different from Google Prediction API? https://developers.google.com/prediction/

It uses a different algorithm. Google Predict likely uses a battery of several types of supervised learning algorithms and lets them "vote". Grok uses their Hierarchical Temporal Memory algorithm.

So you say Google is using a battery of hidden markov chains with classical or slightly novel algorithms?

I could believe that, given the quality of the Google Predict results. Plus Google buying http://recordedfuture.com undermines that their Prediction algo's aren't top notch (mildly expressed).

Re: The Grok prediction engine from Numenta announced

#24

Post by srconstantin »

Element_ wrote: post #22
reader5000 wrote: post #13

I'll remain skeptical until it is determined how their "cortical algorithms" benchmark compared to, say, Naive Bayes. Good marketing though.

Last year Dr. Davide Maltoni published a paper (Pattern Recognition by Hierarchical Temporal Memory)[1] comparing the algorithms laid out in Dileep Georges original PHD thesis. The results were pretty impressive. There have been many improvements to the algorithm since then too. http://bias.csr.unibo.it/maltoni/HTM_TR_v1.0.pdf

Read the paper; HTM's don't seem to do better than other object recognition algorithms at recognizing shapes, especially because there are visual properties it ignores (curvature, global topological properties, etc.) The accuracy for the picture datasets are only 60-70%. What's interesting about HTM is its generality. I can't judge whether it would be good for the Grok prediction engine, but I know more about image recognition and you definitely don't want to use it for that.

Re: The Grok prediction engine from Numenta announced

#25

Post by srconstantin »

sown wrote: post #9

So some ML/AI researchers I know consider Numenta to be of a somewhat "outsider artist" effort. Their opinions of the research from Numenta is that of lower regard. Now, you don't have to upvote me for this, but you don't have to downvote me, either, because I happen to know experts in ML/AI who disagree with Jeff Hawkins. Got that? Downvoting me doesn't change these other people's opinion. I'm just relating to you w…

I don't think "snake oil" is the right paradigm here. In ML/AI, lots of honest researchers are wrong; being a scientist who's wrong doesn't make you a criminal.

That said: Hawkins' principles are very different both from what the brain does, and from what the state of the art in machine learning does. My impression is that HTM's attempt to be too general and assume too little about the problem.

For vision in particular, most successful computer vision algorithms (as well as what we know about the visual cortex's mechanisms) make extensive use of information related to the fact that the image is an image. That is: edges are probably more likely to be continuous than broken; locally constant curvature is more likely than not; textures and colors usually continue over the surface of an object; objects occlude other objects; etc. Brains and effective computer vision algorithms hard-code a lot of information about the nature of the problem they're solving. Hawkins wants to bypass that, and I think it's probably too ambitious an aspiration.

Then again, if he makes it, more power to him.

I don't think we should be prejudiced against someone who comes from the tech industry and wrote a popular book. It's certainly not "snake oil" -- it seems to be a good-faith attempt to solve an important problem. I think the odds are against it working, but that's not a moral condemnation.

Post reply on HN