When an artificial intelligence sees a picture of the comic book hero Spider-Man, it can conjure the idea of a spider 鈥� an advanced connection that humans make and that now seems to be happening in neural networks designed to mimic our brains.
In the 1960s, neuroscientist Jerry Lettvin proposed that human brains had 鈥済randmother cells鈥� that correspond to memories of specific people or objects. The concept was labelled simplistic at the time, but experiments since then using brain implants have shown not only do we have specific neurons triggered by the sight of our grandmother, we also have others that react to Halle Berry and Jennifer Aniston.
Researchers at OpenAI in California have now discovered that a similar system is at work in artificial neural networks as well, suggesting that AIs designed to mimic the function of the human brain are archiving memories and attaching understanding akin to the way we do.
Advertisement
The company鈥檚 CLIP neural network was designed to connect text and images, so the software could find pictures of 鈥測ellow cars鈥� or 鈥済reen apples鈥� based on text input. In experiments, the neural network is given an image 鈥� say, of an old man 鈥� and asked to identify what is in it. Then the images are tweaked and it tries again until it gets to the closest match it can.
While doing this, the software effectively drew pictures of its understanding of different concepts 鈥� so it drew an old man, but it also wrote the words 鈥渙ld man鈥�. Instead of simply matching images to previous images it had seen, this was evidence that it had understanding of the concept behind the words.
The inclusion of the text description itself led the researchers to believe that the AI had developed a kind of grandmother cell related to that concept.
In further experiments, the team found that CLIP has a 鈥淪pider-Man鈥� neuron that fires when shown an image of a spider, the word 鈥渟pider鈥� or even a sketch of the comic book character. Rather than just identifying images that looked similar to pictures of Spider-Man it had seen in the past, it instead seemed to be aware of the concept of Spider-Man, including the word and its association with spiders. They found similar connections relating to geographical regions, facial expressions and religious iconography.
The team believes that these artificial grandmother cells, known as multimodal neurons in the human brain, are a common mechanism in artificial and natural vision systems and that this abstraction of concepts is key to how memory works.
Gabriel Goh at OpenAI believes the experiment shows that neural networks are using the same methods to organise information as human brains. 鈥淲hether you call that understanding or not is maybe a matter of semantics,鈥� he says. 鈥淏ut this is the closest I鈥檝e seen these two systems converge.鈥�
Ilya Sutskever at OpenAI says that the experiments add to evidence of commonality between biological systems and artificial systems, and that neuroscientists will be able to use artificial neural networks to identify behaviours and study human brains for the existence of the same behaviours.
鈥淓arly AI researchers looked at the brain and said 鈥楬ey, the brain has neurons, let鈥檚 try to build neural networks鈥�. It鈥檚 kind of neat that now we鈥檝e reached the point where deep learning can feed back to neuroscience,鈥� he says.
Usually, neural networks for identifying or categorising images need to be trained on large data sets of carefully chosen images and accompanying descriptions. But CLIP was trained on a wide range of different images taken from the internet. Researchers found several examples of racial bias and other problematic associations that had been inherited from the internet. A neuron relating to 鈥淢iddle East鈥� was associated with terrorism, for instance, while another for 鈥渋mmigration鈥� was found to respond to images of Latin America.
But the team believes that small, targeted, educational data sets will be able to decouple these associations and effectively teach the algorithm not to make potentially harmful associations.
Distill
