Artificial intelligence built by Facebook has learned to classify images from 1 billion Instagram photos. The AI used a different learning technique to a great many other similar algorithms, relying less on input from humans. The team behind it says the AI learns in a more common sense way.
Conventionally, computer vision systems are trained to recognize specific things, for instance a cat or a dog. They accomplish that by learning from a huge collection of images that contain been annotated to spell it out what’s in them. After doing this enough, the AI may then identify the same things in new images, for example, spotting a dog in an image it hasn’t seen before.
This process works well, but should be done afresh with every new thing the AI must identify, otherwise performance can drop.
In comparison, the approach utilized by Facebook is a method called self-supervised learning, in which the images don’t come with annotations. Instead, the AI first learns merely to identify dissimilarities between images. Once with the ability to do that, it sees a little number of annotated images to complement the names with the characteristics it has recently identified.
Read more: Demis Hassabis interview: Our AI will unlock secrets of how life works
“The goal was to see if it was possible to make self-supervised systems are better than supervised systems in real scenarios,” says Armand Joulin at Facebook AI Research.
Training the AI took around per month, using 500 professional chips called graphics processing units. It achieved an accuracy of 84.2 % in identifying the contents of 13,000 images it had never seen from the ImageNet database of images, which is often used to classify the potency of computer vision tools.
Joulin says that self-supervised learning is a step towards “common sense” understanding by AI. “It must be in a position to understand anything about the image it really is provided,” he says.
By taking this approach, he and his colleagues think AIs could have a more holistic knowledge of what is in virtually any image. However, the approach requires a large amount of data. Joulin says you will need around 100 times more images to attain the same degree of accuracy with a self-supervised system than you do with one that gets the images annotated.
Read more: Insect brains will teach us learning to make truly intelligent robots
“I would take with a pinch of salt the declare that self-supervised learning alone can lead us to machines which have good sense understanding,” says Nikita Aggarwal at the Oxford Internet Institute, UK. “There’s a difference between developing AI systems that may identify correlations in data to classify images, and systems that can actually understand this is and context of what they’re doing, or indeed reason about any of it.”
Aggarwal can be worried about using images from Instagram to teach AIs to understand about the world. The images will “disproportionately represent younger demographics and the ones who have access to the internet and cell phones”, she says. “There is absolutely no guarantee that this computer vision model will yield accurate results for groups that are not well-represented by the image data set which it has been trained.”
Joulin says that the system hasn’t yet been tested enough to comprehend its biases, nonetheless it “is something we want to investigate later on”. He also hopes to expand the database of just one 1 billion images to further expand the AI’s understanding. “Here we’ve only scratched the surface,” he says.
Article amended on
5 March 2021
We have amended a few of the reported speech from Nikita Aggarwal.
More on these topics:
- artificial intelligence