Brackets picked by artificial intelligence also busted, but see reboot for Sweet 16

The NCAA says there are no verifiably perfect bracket remaining
Published: Mar. 25, 2021 at 6:39 PM MDT
Email This Link
Share on Pinterest
Share on LinkedIn

RAPID CITY, S.D. (KEVN) - It’s safe to say March Madness brackets are busted-- the Big 10 choked with Iowa, Illinois, and Ohio State losing early, the Pac 12 winning big, the South and Midwest regions went up in flames,and many Texas teams had good wins, or even upsets.

Last week, we told you about the South Dakota School of Mines students who used artificial intelligence to pick their bracket for a competition.

Trevor Bormann and Jackson Cates, computer science students at Mines, are competing in a data analysis tournament, called the MinneMUDAC against other Midwest universities.

So far, they are in about 17th place out of 50.

Cates said they have felt the effect of this year’s bracket busting games. But previously, they have been pretty good at picking winners.

Their first round was about 68% accurate, but their second round went up in flames. Cates said he is running the current model on past tournaments to check its effectiveness.

“Some cross-validation techniques, so running our model for previous years to see how well it does and historically, our model predicts like 70% of the bracket correct, but for some reason this year, it’s doing between 50-60%,” said Cates.

Bormann and Cates are now working on predicting scores and weighing blowouts versus close games.

Bormann explains their neural network is like a black box-- they can look at inputs and outputs-- but they do not know exactly what the model is doing on the inside.

“These neural nets are initialized to a random state,” said Bormann. “So it’s all about connections between these hypothetical neurons in the computer. These connections will start out being a random strength, so some will be really strong, and others will be really weak. And then, when we train the model, we are adjusting those weights until the model converges. But depending upon the state it started in, it can converge to different spots.”

The MinneMUDAC Tournament allows for them to re-submit and tweak their brackets for the Sweet 16 games. Bormann and Cates says they did not account for injuries, or COVID-19.

There is a way to reasonably interject something like human error into the model, though it is random.

“It’s exactly human error,” said Dr. Randy Hoover, a computer science and engineering professor at South Dakota Mines, and advisor to Bormann and Cates. “The only difference between human error is we’re going to pick upsets based on favorite mascot or team colors or something like that, whereas this is going to say, ‘randomly select one and pick the loser.’”

The pair said they have been test-running their models, and they said the Gonzaga Bulldogs are consistently chosen as the winner.

Copyright 2021 KEVN. All rights reserved.