Deep learning can seem like magic. But behind that magic there’s often drudgery — painstakingly labeled data is key to many deep learning projects.
“It’s something that has slowed down our industry because the reliance on human annotated data … it’s physically slow,” says Matt Scott, co-founder and CTO of Malong Technologies in a conversation with AI Podcast host Noah Kravitz.
New techniques promise to free humans from this task, and set deep learning loose on a far broader set of problems — like automatically recognizing the products in our shopping carts — Scott explains.
Founded in 2014, Malong — based in China — developed an AI algorithm that learns from the web’s own data, allowing it to classify images that are both “noisy” and unlabeled.
“That’s sort of the context of the problem space we’re working in — how can we access this large-scale data that exists on there on the web, for example,” Scott said.
Earlier this year, Scott and his team entered the WebVision challenge hosted at the CVPR computer vision conference. Competing against more than 100 companies and academic labs, Malong emerged victorious, achieving a 94.78 percent recognition rate.
By contrast, human recognition clocks in at 95 percent.
The platform Malong has built on this technology, Product AI, can take a small photo of a product, and then apply deep learning to parse through images and identify what it’s looking at.
“To get to the next level, we’re going to have to break past the barrier of human annotation,” said Scott. “Now we are in this new area where we are not limited.”
Get Inked with AI: How AI Helps You Choose Your Next Tattoo
Getting a new tattoo? Then give our last episode of our AI Podcast a listen.
We spoke to Goran Vuksic and Dennis Micky Jensen, developers at Copenhagen-based startup Tattoodo, to learn how they developed a neural network that can help you choose the tattoo design for you.
How to Tune in to the AI Podcast
If your favorite isn’t listed here, email us at firstname.lastname@example.org.
Featured image credit: Jayel Aheram, via Flickr.