What is Numenta?

We are creating a platform that allows people to create applications that we think will be intelligent, more in the way human is intelligent than the way computers traditionally have been. Using software tools and arun-time, we are trying to pattern/mimic the algorithms of the brain, and take those algorithms to create new kind of applications.

HTM is a broad theory of the brain and the architecture of the neocortex. In other words, it's a new type of memory system rather than a new type of computing system, that is structured the same way the neocortex is. It's a network structure that you can feed in with sensory data and get an output from it. One day we expect it will be improved by some hardware formation but today it's a software only platform.

Pattern recognition in very large data-sets. Like visual image identification: no computer today can search images, you're searching the text attached to images not the images. Or oil and gas exploration: again a lot of data they need to find patterns in. Pharmaceutical companies are also looking at this: they have a lot of drug testing data, trying to find where the patterns are. Automotive manufacturers have looked at: can the car understand its environment better? That's complex patterns around it, where is it, what is the dangerous situation what's not dangerous. So we see lots of areas where you have lots of data that is difficult to find the patterns.

Why is HTM better than the traditional way computers work today?

Computers are very brittle. There are not very flexible and the human brain is very flexible. So if a dog walked across the road right now, you know it's a dog very easily. You know you never saw that dog before. It's a novel image to you. It's brand new but you would recognise it and you would have no problem with it. In fact, a 2 year old human has no problem with saying that's a dog. But there's no computer today that can look at a dog and say it's a dog. And if you think about it, it's very hard to find the rules of a dog. What is a dog? Four legs and a tail and ears. First of all, that describes a lot of things, not just a dog. But secondly, if you see a dog's front two legs and not back two legs. You would still know it was a dog, no problem.

We are mimicking in this software a hierarchical model of the world. It's a hierarchy in space as well as in time and it creates a model of the world based on the sensory data that it is fed to it over a long period of time. The system has to learn directly from sensory data the way you had to learn directly from sensory data. I mean you learned to drive a car by experiencing it and not by somebody just talking to you about it. So our system needs to learn by exposure of sensory data of the world that it needs to model.

We won't be able to help people with problems where customers can't create a data-set of historical data e.g. here's a hart with no problem, here's another hart with no problem, and another and another, but here's one with a problem, etc... so we have to show lots and lots of harts just like a child is shown lots and lots of dogs till it can create its model that says "I understand what a dog is, I understand what a good hart is".

We plan to license the technology ultimately and provide support, consulting services. Today, we're making it available for free because it's very early. We don't think it's really commercially deploy-able today, it's really in the research phase. Today we have 8 companies signed up to our partner programme: three are involved with national defence issues, one is an automobile company, one designs power plaints, 2 are games company and another is a start-up company.

Nobody has really put it together in the same way we have, but many of the pieces have been available in many research institutions and enterprises. We put it together in a very unique way but it doesn't mean to say that other people won't be trying to do the same. We think it's the next big step in computing and there's going to be a lot of people working on different ways to do the same thing. One way we can differentiate a little bit is with our NuPIC (Numenta Platform for Intelligent Computing) implementation of HTM.

We have a demo today that you can download from our website. But no commercial applications before a year. When i look back at my careers, I really believe that each thing we've worked on has taken almost 10 years to really happen: the Palm Pilot till it became something significant took few iterations/generations; the Treo, it feels it's almost the same. That's why we're funding Numenta ourselves because we don't want to be rushed to it. We want to be patient because we're convinced this is going to be huge and we want to be there.