Daniel Spisak on Technology

When one goes to a conference targeted at computer security professionals it stands to reason that one will hear a lot of keynote speeches about trends within the security industry and where it thinks it is going. At this years 2008 RSA Conference this trend held true with two notable exceptions. First was Al Gore’s keynote on global warming and technology, which the press covering the show was banned from attending. The second was a far more interesting keynote given that had wide ranging implications and very little to do directly with computer security. Jeff Hawkins, the founder of Palm and Handspring and now founder & CEO of his newest company Numenta, gave that keynote. Image: Jeff Hawkins gives his RSA 2008 keynote D. Spisak

In his talk, Hawkins went over problems that Artificial Intelligence (AI) researchers have struggled with, and through trials and tribulations, had grudgingly concluded were if not impossible to solve, extremely difficult at best. One of the problem domains Hawkins examined was visual processing. For example, if you were presented with a series of photos of dogs and cats and asked to classify the animal in each photo, you could do it easily, but a computer using traditional AI techniques would fail miserably.

Hawkins then took a quick tangent through the neural mapping of how monkeys see as a way to introduce the audience to the concept of hierarchical temporal memories. Think of a Hierarchical Temporal Memory (HTM) as a pyramid; at the lowest level you have the greatest number of neurons processing visual input in a very basic way and passing up the hierarchy a reduced set of information that encapsulates the input seen. The other key aspect of this system is the temporal aspect of input processing; we don’t just see what is happening in the present—we have a memory of what things looked like in the recent past as well, which helps us process the massive amount of visual data coming in every second. This process happens at each level: at each step the amount of information and speed of its flow is reduced. Crucially, information can pass bi-directionally between different levels of the hierarchy. It is this bidirectional communication that allows you to see a photo of a cat, recognize the shape, color, and other aspects passing them along to your brain, which then processes this information and recognize the photo as a cat.

Now, why is all of this important? With this enhanced understanding of how information is actually processed it’s possible to try and build an electronic model of an HTM. This is where Hawkins’ company, Numenta, comes into play, creating an electronic HTM development platform for Windows, Mac and Linux platforms called the Numenta Platform for Intelligent Computing or NuPIC. (You can download it here.) NuPIC has its own language and syntax, affording developers near-limitless flexibility to build HTM based applications. To help users new to this development platform see what is possible, Numenta has created a demonstration HTM application called Pictures. This HTM application takes in as input a set of photos scanned in at multiple angles and then builds up an ability to start recognizing these photos, even when the images are obscured with noise.

In Hawkins’ talk, he showed recognition results for the Pictures HTM application under different kinds of visual impairment (i.e., static noise, dynamic noise, etc) that were surprisingly good. As the amount of visual noise became denser and more complex the HTM’s accuracy did decline, but was still significantly useful. According to Hawkins, based on Numenta’s experience with creating and enhancing the Picture’s HTM application he predicts that the computer vision problem of asking a computer to ask if a picture is of a dog or a cat could be solved within the next two years. Hawkins also showed examples of photos the Pictures HTM recognized as certain types of objects, like cars or boats when the application had been trained with photos of those objects and then given a set of random photos to try and determine what they are. This demonstration of the power of HTMs is but scratching the surface of what is possible.

Hawkins is keen to caution that HTMs are not a cure-all solution to all kinds of AI/computer learning domain problems. HTMs assume your problem domain can be expressed as a spatial-temporal hierarchy. This is a technology that will have far-ranging implications once it starts to get into the hands of the truly imaginative and creative developers out there. While Hawkins technology currently doesn’t have immediate implications within the security realm, it is easy to see how one might apply the technology to age-old problems within the industry. Natural applications for an HTM could be processing log data from IDS and firewall systems, better face recognition systems, ability to identify dangerous objects in photos or videos.

These are just a few of the applications that this writer can think up. There are likely far more out there that I haven’t thought of that could be just as important or better. Year after year we hear about new security products and services from the industry aimed at businesses and solving their security challenges, but the process has become an incremental one. This was evident at this year’s RSA show. What is needed is something revolutionary, not evolutionary, to help the industry. Perhaps HTMs will turn out to be something these security companies will add to their arsenal and turn into useful tools and solutions that we need.

(Originally posted at TechRevu)