|
|
Projection of voice onto feature subspaces used by recognition systems
To illustrate that perhaps current voice recognition systems may not be looking at the
right set of features, I performed a simple experiment. Voice signals were projected onto
the feature set used by recognition systems, varying levels of noise was added in dimensions
orthogonal to these features, and then the signal was back-projected. The results are surprising.
|
ImageMosaics
ImageMosaics have recently replaced autostereograms as the
new cool thing in computer vision / graphics. Here are a few examples.
|
|
|
Graphics
Many of the projects which I've done that branch into the realm of computer graphics.
|
Images to ASCII Revisited
I have revisited the old concept of representing images with
standard ASCII characters. I have improved upon the old methods by
not only modeling the gray levels, but also the shape of the
images. Using this same idea, I have also designed a converter which
traces the edges of binary images.
|
|
|
Learning to Play PacMan Using Incremental Reinforcement Learning
I have applied an
"embodied intelligence"
behavior learning approach to
learn optimal playing strategies for both PacMan and the opponent ghosts.
Over many thousands of iterations, the linear creature-centered perceptron
network is continually modified with reinforcements based on rewards
gained through out each game. To make such a formidable learning task
feasible, we begin training with extremely simple boards, then increase
their complexity once simpler, basic behaviors have been learned.
|
Page loaded on April 19, 2024 at 12:25 AM.
Page last modified on
2008-04-14
|
Copyright © 1997-2024, Jeremy S. De Bonet.
All rights reserved.
|
|
