LizardDetector

Moral of the story (to save you some reading): Neural networks are ALMOST NEVER the correct solution for a problem.

• Actual moral: Use a variety of training data in any supervised machine-learning task. Because supervised machine-learning (~= a neural network) is almost certainly the correct method for automated lizard-detection. --MikeBuchanan

MorganConbere (the immortal savior of the world) took Neural Networks with ProfessorKeller in the fall of 2006. The class ended with a final project or paper. Not being one to volunteer to write papers, Morgan decided to make a project. Building off of his summer research, he decided to write a lizard detector.

The problem: Given a grainy image of sand and gravel and lizards, identify the lizards in the image with x,y coordinates.

The solution: Morgan wrote a neural network framework in Java to process the data. Before he could test it he needed data, so he wrote a java applet that could be used on a website to enter the data. He also wrote some CGI scripts to take the result of the applet and place it into a nice format that the framework could read. All of this framework took him nearly up until the deadline of the project. Hours before the deadline, he finally was able to put all of this data together and tested the program. Much to his surprise, it detected every lizard!

The problem: It detected every lizard. It detected every not-lizard. It detected everything. Morgan had written a neural network that learned to find lizards. All of his training data was of actual lizards. There was no data that said to find not-lizards, so the neural net learned very quickly to return true on any input. Morgan gave a sleep-depped presentation about this to his class and promptly never thought about writing neural nets again.

Praise for Morgan's Lizard Detector: "A scientific instrument with perfect sensitivity but really lousy selectivity" -- PhilMiller