Facial Recognition Mistakes: Last Week’s Clearview AI Episode

Play Audio Clip Listen to audio clip.

This week’s “Fox Nation: Clearview AI” makes a new project that worked better than it should have done: After a new study highlighted a facial recognition system from C4 Tech as a terrible one, it decided to do something different.

It’s something we first saw in last week’s episode of “Fox Nation: Clearview AI”.

“The key factor in facial recognition is noise,” said Sarah. “When the system makes a match and the data points comes from a benign noise, such as a train or a bike, it will make the match incorrectly.”

Sarah worked on the project last year where she and Anjan Sundaram, executive director of Clearview AI, did their own study of the system and found they made the wrong match every two out of five times. She asked Sundaram to try another approach and it had a much better number.

“The reason why we did the regression and not a main test like in the C4 Tech study was it was completely valid for our test because we were looking at the best data points that we have available to us,” said Sarah.

After first trying the C4 Tech system and finding it to be very poor, she and Sundaram decided to use C4 Tech’s main test which is 10 data points to see how well it would do on their new database.

They added 60 accounts to the database for extra randomness, real photographs and recreations and compared their results to C4 Tech’s data points.

“We’ve found a change in the system. It now works well as originally designed by C4 Tech. The individual difference between their original dataset and ours was insignificant.”

Sarah says now the system is working properly again she hopes its findings will show up in all face recognition systems to improve accuracy.

“If C4 Tech’s original system is at least working fine then there is nothing to be concerned about it. This was just kind of a hiccup.”

Subscribe and subscribe again at foxnation.com

Leave a Comment