Unsupervised neural networks will never match models trained on well-labeled data
Matthew Zeiler · Matthew Zeiler, Clarifai // Data Driven #31 // Nov 2014 (Hosted by FirstMark Capital) · Nov 20, 2014 · at 15:48
Matthew Zeiler (founder of Clarifai) answers a question from Matt Turck (FirstMark Capital) about training neural networks for vertical domains like medical imaging.
0:00 / 0:10exact quote · 10.2s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“These models work really well with labeled data. There are approaches where you can train without any labels but the performance is never as good as if you have well-labeled data.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Matthew Zeiler
Clarifai does not view Google as a direct competitor
“So Google, we don't consider them a competitor. They have great research teams and really big research teams and great resources but they're working on their own problems. They have their own users, and they don't they don't really compete with us in terms of …”
Competitiveness in computer vision now strictly requires using neural networks
“And you can see now, to even be competitive, you have to use neural networks.”
Clarifai won the seminal 2013 ImageNet computer vision competition
“Clarify happened to be this result at the bottom. We won the 2013 competition.”
Clarifai's video recognition runs ten times faster than real time
“It can run about 10 times faster than real time,”