Monday, March 10, 2008
If you took photographs of a mountain for a year about 20 times a day and recorded the weather and time for each shot, then you could use machine learning to learn what the mountain should look like as a function of the date and time.
If you did that using 3 perceptrons per pixel to learn linear functions for hue, color and saturation for each pixel, then you'd get this image for May 20 at 2:20 pm. Not quite convincing yet, but not bad for such a simple model.
That's in 10 days. And for it to be correct for this year, a lot of snow better melt!
Algorithm and image by CS 658 student Ilya Raykhel who is a graduate student in machine learning.
Posted by Mike Jones at 1:46 PM