It is very difficult to create an AI that can perform as well as the human brain. A new study by researchers at the Imperial College London, published on October 4th in Nature Communications, explores one reason why AI underperforms – their perfection.
“The most striking difference between humans and AI is that our learning seems to be much more robust,” senior author Dr. Dan Goodman told me. Humans are fantastic learners. Even babies can quickly learn to identify images or can generalize their learning in one area and apply it to another. “If you train them [AI] with one dataset and then test them on slightly different data they often fail spectacularly, whereas people do much better on this.”
“Humans can [also] learn with much less data than AI systems,” Goodman continues. “I can learn to recognize a new animal from just one or two pictures, whereas an AI system probably needs hundreds or thousands of examples.”
In a neural network that makes up an AI, each cell is identical. In a human or animal brain, however, each neuron is unique. “They differ in hundreds of different parameters, including their electrical and chemical properties, their shapes, etc,” says Goodman.