Thinking as a Hobby

Get Email Updates
Email Me

Admin Password

Remember Me

3477905 Curiosities served
Share on Facebook

AI Research
Previous Entry :: Next Entry

Read/Post Comments (3)

I spend most of the time talking political, but I do other things in addition to ranting.

Sometimes I give updates on the independent AI research project Philip and I are working on, and it's been a while, so if anyone's interested, here you go.

Our opensource implementation of NEAT, called ANJI, is still being actively maintained.

We've got two journal papers in submission and under review. One on simplification/complexification dynamics in the evolution of artificial neural networks, and the second for a neural network-controlled active vision system for identifying simple shapes.

The idea is that we're evolving a "roving eye", and agent that scans a 2D image by moving around on its surface, left/right, zooming in/out, and rotating. The agent is controlled by an artificial neural network, a biologically-inspired bit of software that works like a very simple biological neural network.

In our last round of experiments, we tried to evolve eyes that could distinguish between simple shapes: squares, circles, and triangles. Our results in the submitted paper had identification rates above the 90% range, but we've since improved to where our eyes can identify the shapes correctly 100% of the time.

I don't know if this link will work right here, but here's a video of the same eye scanning a series of six shapes: 2 circles, 2 squares, and 2 triangles (Flash plug-in required).

The upper left part of the screen shows the eye moving around the canvas. The upper right shows the eye's point of view, and the lower portion of the screen shows the affinity output of the network...basically its confidence that the shape is the one being identified, in this case a triangle. So the affinity should stabilize on 0 when the eye sees circles and squares, and should stabilize on 1 when viewing a triangle (which it does).

So, we're now working on a much more difficult domain, applying the same approach to fingerprint classification, which basically involves the initial sorting of fingerprints in an identification database into one of 5 classes: left loop, right loop, whorl, arch, and tented arch. If you look at your fingerprints, they should all roughly fall into one of these classes (a very small percentage are unclassifiable).

We're hoping to complete this work by mid-January and present our findings at the GECCO conference in 2005 in Washington DC.

Read/Post Comments (3)

Previous Entry :: Next Entry

Back to Top

Powered by JournalScape © 2001-2010 All rights reserved.
All content rights reserved by the author.