The Prediction Model

I've become increasingly interested in the overlap of biology and computing over the past few years. It began with the realisation that the web and stock market are really biologic in nature with their fault tolerance, nodal shape, replication of information and distributed locus of control and was further prompted by work I undertook on AI systems for Advertising and Social Computing solutions.

I’m enjoying this overlap developing into a moderate obsession and I am trying to steer my thinking on all things computing into a more ‘biologic fashion’. I’ve always been a strong believer that people involved in one discipline can offer fresh insights on other sciences and that a good set of ‘first principals’ can work well cross domain. This cross pollination was the grease that helped the machine of the Industrial Revolution into being and obliquely it’s also the reason I give for sporting sideburns like some Manchester factory owner.

This post is inspired by Jeff Hawkins who is doing work into models of the brain and attempting to derive an overarching theory of the brain which is something that, despite the reams of data we have on the brain, we are as yet unable to articulate. His talk was on the use of a Prediction Model as the primary approach to developing a theory of the brain and he got my mind racing.

After graduating from Cornell in June 1979 he read a special issue of Scientific American on the brain. In it Francis Crick lamented the lack of a grand theory explaining how the brain functions.^[3] Initially, he attempted to start a new department on the subject at his employer Intel, but was refused. He also unsuccessfully attempted to join the MIT AI Lab. He eventually decided he would try to find success in the computer industry and then try to use it to support his serious work on brains, as described in his book On Intelligence

Jeff thinks that the reason we still haven’t managed to define intelligence well is that we don’t have this overarching theory of the brain or more accurately – intelligence. Jeff postulates that the brain isn't like a powerful computer processor and that instead it’s more like a memory system that records everything we experience and helps us predict, intelligently, what will happen next.

Things like these stop me sleeping at night and last Sunday I leaned over to my girlfriend at 2am and whispered to her “I have to write some stuff”.

I slipped out of bed and knocked up the notes below. They are presented here un-edited and what you see is the first pass brain dump of some of my thoughts and concepts surrounding a Prediction Model (It's probably best to click on one and open up the set in Flickr and view from there).

If you are involved in this area at all I would love to hear from you as I intend to delve deeper. Physics has alot to add to this area with work in quantum theory and calculations surrounding boundaries of event horizons for black holes all being of relevance to the model of the brain and prediction.

Stanford's Make3D - 3D flythroughs from a single 2D image

Make3D is a project from two Stanford students that won a 'best paper' award at the International Conference on Computer Vision in Rio de Janeiro in October 2007.

The online service takes a 2D image and creates a 3D'esque fly around models that include depth and a range views. Photos can be uploaded directly or pulled into the site from Flickr - however the service requires that you rate at least five images before you can pull them in from Flickr. You can jump in and just use the upload from hard drive option though

Click through to this link to see an example of it in action from a photo I took in Egypt at the pyramids:



Note: it doesn't work for Intel Mac and Shockwave but there is a workaround for installing a VRML viewer for Linux which will be given to you as an option.

The tech lowdown on the algorithm from January Stanford News Service :

…the algorithm breaks the image up into tiny planes called “superpixels,” which are within the image and have very uniform color, brightness and other attributes. By looking at a superpixel in concert with its neighbors, analyzing changes such as gradations of texture, the algorithm makes a judgment about how far it is from the viewer and what its orientation in space is. Unlike some previous algorithms, the Stanford one can account for planes at any angle, not just horizontal or vertical. This allows it to create models for scenes that have planes at many orientations, such as the curved branches of trees or the slopes of mountains.

An excellent diagram on the process can be found here.

This is one of the first web services of this ilk out of the starting gate. Microsoft have had Photosynth for a while but instead of using one image to derive the model it meshes together multiple images.