Artificial intelligence can do remarkable things, like recognize faces on social networks, instantly translate speech from one language to another, and identify commands barked into a smartphone. But it also can do stupid things, like label an African-American couple “gorillas.”
The artificial intelligence underpinning Google Photos did just that last year. The platform uses deep neural networks to identify images in your photo collection. These networks of hardware and software, modeled after the network of neurons in your brain, learn to recognize objects, animals, and faces by analyzing many millions of pre-labeled photos. It works incredibly well, but as Google proved, it’s not perfect. And so the company decided to stop labeling anything as a gorilla. (And apologize profusely).
Researchers strive to solve the sometimes egregious limitations of this breed of AI, called deep learning, as it evolves. Matthew Zeiler, the founder and CEO of the New York startup Clarifai, is developing deep learning technologies similar to Google’s. He’s offering them to the world’s businesses to use as they like. And he’s offering tools that he hopes will allow them to sidestep the kind of gaffe Google experienced with Photos.
It’s part of a broader effort to democratize the deep learning technologies created by the likes of Google, Facebook, and Microsoft. Companies like Algorithmia and MetaMind (now owned by Salesforce.com) offers services similar to those provided by Clarifai. There’s an online marketplace for deep learning algorithms. And even Google and Microsoft are beginning to offer deep learning APIs to outside businesses via their computing services.
When it launched in 2013, Clarifai would train deep learning models for customers. Now it lets them train neural nets of their own. That may sound daunting, but the company hopes to ease the process through a simplified user interface. Zeiler says you can train its image recognition system on as few as 10 data examples with no coding necessary. You can refine the parameters with more manual controls. You can train an AI model to recognize shoes, for instance, and then, by tagging a few Nike shoes, you can teach it recognize Nikes.
Businesses could use this for e-commerce. They could allow customers to snap a photo of a piece of furniture, upload it to a website, and see who makes it. Businesses could also use the system to filter unwanted content like nudity from their sites. By democratizing the training of deep learning, Zeiler says, the system can avoid the situation like the gorilla gaffe. “To solve some of the gaffes we’ve seen, we need a diverse set of users,” he says. “We need them from different backgrounds and different viewpoints.”
Independent AI developer Guarav Oberoi is skeptical. According to him, any AI model is going to get some predictions wrong. But hopefully, as time goes on, the people training AI will keep this to a minimum.
This article was syndicated from wired.com