for a decade now, many of the most impressive AI systems have been taught using a huge inventory of labeled data. An image can be labeled “tabby cat” or “tiger cat”, for example, to “train” an artificial neural network to correctly distinguish a tabby cat from a tiger. The strategy has been spectacularly successful and woefully flawed.
Such “supervised” training requires painstakingly labeled data by humans, and neural networks often take shortcuts, learning to associate the labels with minimal and sometimes superficial information. For example, a neural network could use the presence of grass to recognize a photo of a cow, because cows are often photographed in fields.
“We are raising a generation of algorithms that are like college students [who] they didn’t come to class all semester and then the night before the final, they’re overcrowding,” he said. Alexei Efros, a computer scientist at the University of California, Berkeley. “They don’t really learn the material, but they do well on the test.”
Furthermore, for researchers interested in the intersection of animal and machine intelligence, this “supervised learning” might be limited in what it can reveal about biological brains. Animals, including humans, do not use labeled data sets to learn. For the most part, they explore the environment on their own, and in doing so gain a rich and robust understanding of the world.
Now, some computational neuroscientists have begun to explore neural networks that have been trained on little or no human-labeled data. These “self-supervised learning” algorithms have been enormously successful in human language modeling and, more recently, image recognition. In recent work, computational models of mammalian visual and auditory systems built using self-supervised learning models have shown a closer correspondence with brain function than their supervised learning counterparts. To some neuroscientists, it appears that artificial networks are beginning to reveal some of the actual methods our brains use to learn.
Brain models inspired by artificial neural networks came of age about 10 years ago, around the same time as a neural network called AlexNet revolutionized the task of classifying unknown images. That network, like all neural networks, was made of layers of artificial neurons, computational units that form connections with each other that can vary in strength or “weight.” If a neural network can’t classify an image correctly, the learning algorithm updates the weights of the connections between the neurons to make that misclassification less likely in the next round of training. The algorithm repeats this process many times with all the training images, changing the weights, until the network error rate is acceptably low.
At about the same time, neuroscientists developed the first computational models of primate visual system, using neural networks like AlexNet and its successors. The match seemed promising: When monkeys and artificial neural networks were shown the same images, for example, the activity of real neurons and artificial neurons showed an intriguing correspondence. Artificial models of hearing and odor detection followed.
But as the field advanced, researchers realized the limitations of supervised training. For example, in 2017, Leon Gatys, a computer scientist at the University of Tübingen in Germany, and his colleagues took an image of a Ford Model T, then superimposed a leopard skin pattern on the photo, generating a strange but interesting image. easily recognizable. . A leading artificial neural network correctly classified the original image as a Model T, but considered the modified image to be a leopard. He had been fixated on texture and didn’t understand the shape of a car (or a leopard, for that matter).
Self-supervised learning strategies are designed to avoid such problems. In this approach, humans do not label the data. Rather, “the labels come from the data itself,” he said. Friedemann Zenke, a computational neuroscientist at the Friedrich Miescher Institute for Biomedical Research in Basel, Switzerland. Self-monitoring algorithms essentially create gaps in the data and ask the neural network to fill in the blanks. In the so-called large language model, for example, the training algorithm will show the neural network the first few words of a sentence and ask it to predict the next word. When trained on a massive corpus of text pulled from the Internet, the model seems to learn the syntactic structure of the language, demonstrating impressive linguistic ability, all without external tags or supervision.