The way that babies learn about and navigate the world could prove to be a good model for training artificial intelligence.
AIs don鈥檛 learn as efficiently or flexibly as children. To explore why, Brenden Lake and his colleagues at New York University turned to the SAYCam data set, which was published this year. It contains video footage from head-mounted cameras worn by young children for a few hours each week over their first three years of life.
The team fed an AI neural network raw data from SAYCam and asked it to try to work out what it was seeing by identifying what was unchanged and what was different in the video stream over time.
Advertisement
The AI did begin to make sense of the videos. For instance, it was able to recognise that the same object 鈥� a cat 鈥� popped up repeatedly in the videos. But the AI often did so by extending its attention beyond the cat itself, suggesting it may be relying on contextual cues to identify objects.
[video_player id=鈥漍3OiSevK鈥� access_level=鈥漞veryone鈥漖
Lake says this suggests the algorithm doesn鈥檛 recognise objects in the same way as a child, but he argues the findings are still significant. 鈥淲e have a proof of concept that [visual features] are learnable with enough naturalistic data,鈥� he says.
Simone Scardapane at Sapienza University, Italy, says the work offers a 鈥渇ascinating insight鈥� into how AI algorithms would react when they are fed the kind of messy data children must deal with rather than the heavily engineered data sets they are normally given.
鈥淎I requires a lot of data and labels in order to get to the same levels of performance on a task that kids are good at,鈥� says Lake. But if it becomes possible to train them to learn in similar ways to a child, their intuition may well be stronger.
Reference:
