Do you trust Elon Musk and Stephen Hawking when they warn that artificial intelligence, particularly in autonomous weapons, may be advancing faster than we can control it? Have you ever been steered wrong by a Google map? Would you trust Google in more difficult tasks, like creating artificial intelligence? Do you believe Google has the best interests of humanity in mind in all that it does? Would you be excited if Google announced it had developed an artificial intelligence that created its own AI child that can outperform humans? Would like to check with Elon and Stephen again? Do you think it’s too late?
Researchers at Google Brain – a name that seems to be becoming more oxymoronic by the day – announced this week that they have developed an artificial intelligence called AutoML, which is short for Automated Machine Learning, but the ‘M’ could also stand for ‘Mother’ because its main purpose is to develop and generate its own artificial intelligences. You could call this new AI a ‘child’ but you’d be too late because Google Brain has already thought of that. However, to reduce the possibility of panic, its official name is the more innocent NASNet.
In a post on the Google Research blog, the researchers explain that AutoML is more than just a parent -- it’s a teacher as well. In the described experiment, AutoML trains its child NASNet to recognize objects in a video – things like people, cars, clothing items, etc. If this sounds like a human parent pointing to pictures in a book and getting their child to say “cat,” you’re right. If you can imagine that human parent correcting the child who said “cow” instead of “cat,” you’ve also described what AutoML does to NASNet. That doesn’t sound so bad, does it?
Oh, you gullible humans. Unlike a human parent, AutoML can correct and repeat this training thousands of times without getting frustrated, hungry or tired, and NASNet can endure this repetitive training without getting fidgety or needing to use the bathroom. Once the education was complete, NASNet was tested on two well-known datasets -- ImageNet for images and COCO for objects – and outperformed all other computer vision systems.
Think about that for a minute. A machine made by a machine outperformed the best machines made by humans. Are we ready for this? Is Google?
“We suspect that the image features learned by NASNet on ImageNet and COCO may be reused for many computer vision applications. Thus, we have open-sourced NASNet for inference on image classification and for object detection in the Slim and Object Detection TensorFlow repositories.”
Open source! Without any standards nor regulations in place, the Google Brain (do you see the oxymoron yet?) has unleashed its AI and its fast-learning child upon the world. Alphabet’s (Google’s parent company) own DeepMind company, which is supposed to be working on issues concerning the moral and ethical development of AI, didn’t have anything to say.
It’s easy to see how advanced object recognition will help applications like driverless cars. Are we ready for driverless cars to begat driverless golf carts that are smarter than their parents? Will they take us where we want us to go … or where they plan to dump us?
Elon? Stephen? Anyone? Bueller?