Project Portfolio

Dan Maynes Aminzade

Research

Actuated Workbench
Audience Interaction
Hover
You're in Control
Edible User Interfaces
Fuzzmail
KC-135
OSCAR

Schools

Stanford
MIT
Carnegie Mellon

Industry

MERL
Microsoft
Adobe
Disney Imagineering

Fun

Unsafe Search
Music Visualization
Mobot
PantsCam
Taboo Database
Pointillism
Painting
WebAmp

Zany

Tacos
SETI Joke
Pepsi Database
Love Calculator

Hacks

AdBall
RCA Lyra
Stone Cold

Humor

SURG Proposals
Female Pop Singers
Satan Baby
Wesley Willis

Taco AI

Taco Bell recently introduced a new menu item called a "chalupa".  The commercials for the chalupa say that it is "crunchy" and "irresistible", but they never really explain what a chalupa is.  Is a chalupa more like a taco or a burrito?  After pondering this important question, Ethan and I decided that a computer could give us the answer.

We drove to the local Taco Bell armed with a digital camera and took a series of photographs of the contents of our value meal.  We captured images of tacos and burritos of all sorts: beef, bean, and chicken, soft shell and hard shell.  We then took a series of photographs of the mysterious new chalupa.

 


We trained the neural net on this battery
of sample images of tacos and burritos.

Using the source images of tacos and burritos, we trained a neural net to distinguish between tacos and burritos.  The neural net does this by establishing a weight for each pixel, summing the weighted differences across a test image, and comparing this sum to a threshold value.  We found that after training our neural net with ten taco images and ten burrito images, it was able to distinguish between new pictures of tacos and burritos with almost perfect accuracy.
Next came the true challenge: the chalupa test.  It turned out that the chalupa images were almost invariably identified as "tacos".  After four hours of programming and ten dollars worth of Mexican food, we had our answer: the new chalupas at Taco Bell are more like tacos then burritos.


In this neural network sensitivity graph,
lighter colored areas represent pixels that
are most heavily weighted when
distinguishing between a taco and a burrito.

Our research didn't stop there, however.  I immediately thought of a new question: am I more like a taco or a burrito?  I took a series of ten images of myself making various facial expressions and we passed them on to the neural net.  Of these ten images, eight were identified as "burrito" and only two as "taco".  Once again, the computer provided an answer: my composition is roughly 80 percent burrito and twenty percent taco.