Numenta Blog

The implementation of these algorithms is relatively straightforward. The challenge is the state of the art in computer vision is not currently at a point where it is possible to reliably detect 10s-100s of object categories in real time on current systems. It is currently possible to build systems that get decent real-time performance on detecting a few categories concurrently, or offline systems that get around 60% accuracy across hundreds of well-defined categories. Thus, (1) faster general purpose hardware, (2) better algorithms, or (3) running the best algorithms in ASICs designed for CV is necessary. The top labs now typically use GPU clusters to train / run their algorithms, with the computationally expensive stages usually being feature extraction and/or classifier training.

Google Predict (http://code.google.com/apis/predict/) offers a general machine learning API geared towards those who want to apply machine learning to their applications without subject-specific knowledge. I've not used it so I can't speak to its accuracy, but it is not geared towards computer vision and I imagine it would fail miserably at such tasks (since computer vision is highly dependent on domain-specific feature extraction techniques), and I imagine it performs well at NLP tasks. The primary limitation of such a system is that it acts as a black box - you throw data in and get answers out without any knowledge of the process behind it.

This black-box model is limiting for three…

I interned at Numenta [0] in 2012. Numenta is building an open-source machine intelligence product [1] based on the human brain. Specifically, the algorithms are based on the theory of "hierarchical temporal memory" (HTM) [2] as described by Numenta founder Jeff Hawkins [3] in his book On Intelligence [4]. The basic idea is that the neocortex has a generalized learning framework that acts on generalized input from all five senses. For example, one study enabled blind people to "see" a ball in front of them and grasp it with their hands, by encoding the image of the ball onto electrical impulses on taste buds, using a device in contact with the tongue. I can't find a link to the original study, but here's an article that describes one such device. [5] The conclusion drawn from this is that, even though blind people have lost their sight, their neocortex is able to understand inputs from any sensory source (in this case, tastebuds). The Numenta algorithms generalize this idea to feed time series data into a virtual "neocortex" that builds custom ML models for each data source and chooses the best one (a form of online learning that eliminates the need for data scientists to create custom models for each data source).

Because the Numenta algorithms are optimized for time-series data, their areas of strength are prediction and anomaly detection. When I worked there (5 years ago, they've made a lot of progress since then), one of their main "case studies" was energy usage in…

Oh don't get me wrong, the things being done with ANNs are incredibly impressive. Far more so than anything I've seen with HTMs to date. I was just pointing out that Jeff is putting his money where his mouth is. Also, much of the resurgence in interest in AI came about 12 years ago in large part due to Jeff's thinking and writings (which was mostly consolidating existing ideas into a cohesive theory), as Andrew Ng and others have stated[1]. Jeff helped kick AI out of a plateau and avoid another AI winter.

For interesting things that Numenta has done, just check their website[2]. Their most popular product is likely Grok[3], which predicts things like server loads. It's nothing that ANNs couldn't do, and ultimately isn't that impressive at face value, but the implementation is interesting and has far reaching implications (w/regards to things like energy efficiency and other areas).

I think the future holds incredible things for ANNs and we'll continue to see great breakthroughs with them, but HTMs hold a lot of promise too. ANNs have been around quite a bit longer and have orders of magnitude more researchers working with them, so I'd argue it's premature to dismiss HTMs just because a handful of researchers in Redwood, CA haven't solved every AI problem within 12 years.

It was only 15-20 years ago that ANNs were considered to have peaked and been surpassed by bayesian networks and support vector machines. How times have changed since then!

[1]…

So I figured it out. Basically, they take the idea of AGI seriously, and actually consider and talk about the repercussions, and therefore you dismiss them and their ideas as fringe and not worth investigating. I know that, because if you had investigated at all, you would see that all of those projects had really interesting results and these people are not being vague and hand-waving.

Not all of those projects I listed identify themselves as AGI. However, they should go in the same group.

And anyway, all of those projects have demonstrated progress. If you looked into them at all then you would see that. Ben Goertzel is using some aspects of his AGI research in mainstream (narrow) AI projects. OpenCog has released a number of solid demonstrations of current features. And Goertzel isn't hand-waving or bullshitting in his numerous books and scientific papers, for example Probabilistic Logic Networks: A Comprehensive Framework for Uncertain Inference (336 pages).

Hawkins has demonstrated very interesting progress with his software and has a commercial application https://www.numenta.com/grok_info.html

Voss is using his system at Adaptive AI as a commercial enterprise.

Qualcomm is funding Brain Corporation (Izhikevich et al) so obviously they are taking it seriously. A bakery in Tokyo has tested Brain Corporation's machine vision technology to power a semi-automated cashier system

http://www.diginfo.tv/v/12-0145-r-en.php

If they had a solution that was better than what was out there, and they are publishing their research, then it follows we should be able to find their novelty in their publications, right? So, I took a look at their publication track record on this topic [0], and there are a few red flags that stood out to me.

First, they're publishing at a pretty astounding rate, with new submissions roughly every 1-3 months for the past 3 years, with the same authors. I'm skeptical that a small (<5?) research team would be able to produce novel contributions at such a rate, which leads me to believe they're probably remixing previous works into incremental publications.

Second, all of their publications are in neuroscience journals, which may not be up to date with the latest research going on with HTM in (for example) computer science. NIPS does have a computational neuroscience track, after all - I would have expected at least some interdisciplinary leakage if the novelty was very high.

Finally, the journals that they're publishing in are relatively low-ranking, which also seems to indicate that what they're doing is not necessary very novel (as you said, perhaps just a re-hash of existing HTM research).

Also, this article was co-written with a marketing exec, so I think it's fair to bring a certain level of skepticism.

[0]: https://numenta.com/resources/papers/

For people wanting to look into HTM (Hierarchical Temporal Memory), do check out Numenta's main website [1], in particular the papers [2] and videos [3] sections.

Otherwise, HTM inventor Jeff Hawkins' book "On Intelligence" is one of the top 3 or so most fascinating books I've ever read. It doesn't cover HTM though, just how the brain works at a conceptual level, but in a way I haven't seen anyone else explain. Jeff clearly has an ability to see the forest through the trees in a way that is not too commonly found. This is one of the reasons I think HTM might be on to something, although it of course has to prove itself in real life too.

But we should remember for how long classic Neural Networks was NOT overly successful, and almost dismissed by a lot of people (including my university teacher who was rather skeptical about them, when I took an ML course on like 12 years ago and personally believed a lot in them). We had to "wait" for years and years until enough people were eventually throwing enough work on finding out how to make them really shine.

[1] https://numenta.org/

[2] https://numenta.com/neuroscience-research/research-publications/papers/

[3] https://www.youtube.com/user/OfficialNumenta

[4] https://www.amazon.com/Intelligence-Understanding-Creation-Intelligent-Machines/dp/0805078533

Edit: Fixed book link.