By Julia Schwarz
Brenden Lake, an expert in learning and intelligence, has joined Princeton as an associate professor of computer science and psychology.
“I’m thrilled to be here,” Lake said. “Princeton has always been at the forefront of research in human and machine intelligence.”
Lake spent eight years on the faculty at New York University, most recently as an associate professor of psychology and data science. He studies intelligence in machines and people, how each learns differently and how they draw on what they know.
“Despite the amazing progress in AI, people are still smarter than machines in so many different ways,” Lake said. “We’re trying to understand why.”
Much of Lake’s recent research has focused on differences in language acquisition between humans and machines. “There are fundamental differences in how AI systems come to have their abilities and how people do,” he said.
Modern AI systems, for example, require a much larger amount of training data than children to become intelligent. GPT-4, for example, was trained on trillions of words. If a child needed to hear as many words as an AI system to understand language, fluency would take at least 100,000 years, Lake said.
Several studies from Lake’s lab have investigated whether machines can learn from the same limited inputs as children, using datasets of headcam video recorded intermittently by young children between the ages of 6 and 25 months. The lab used the data to train AI systems on input from individual children. One study found that a small language model, trained from scratch on language input to just one child, can learn meaningful linguistic structure. Another study used all 61 hours of video and audio to successfully train a neural network from scratch to map words to their visual referents, such as the word “ball” and a picture of a ball.
The results of this research, Lake said, demonstrate the power of experiential data in learning and intelligence: a neural network can indeed extract knowledge from the same data that a child sees and hears by the time they turn two.
But the research also showed that the level of intelligence produced by the same data in both a human and a machine was remarkably different. A neural network trained on two years of a child’s experience is still nowhere near as intelligent as a two-year-old. The machine’s knowledge was “far more fragile, less robust, less intricate,” said Lake.
The reason for this difference in outcome is an open question, he said. It’s likely because children have access to additional experiences — feel, taste, touch — that machines don’t. “My daughter’s first word was banana, because she loved eating bananas,” Lake said. “An AI system couldn’t care less about the taste of a banana.”
Lake’s work is deeply rooted in both psychology and computer science, two disciplines that have long been intertwined. Many early pioneers of neural networks, for example, were psychologists.
“There’s always been a back-and-forth between the fields,” said Lake. There is so much that AI can do to help cognitive scientists, he added, but also so much that cognitive science can give AI in terms of explaining how human minds are able to be both data efficient and flexible.
Lake is also an affiliated faculty at the Princeton Neuroscience Institute and is on the executive committee of the Natural and Artificial Minds initiative, part of the Princeton AI Lab.
Lake’s research has been recognized with a Robert J. Glushko Prize for Outstanding Doctoral Dissertation from the Cognitive Science Society. In 2018 he was named by one of the 25 Innovators Under 25 by the MIT Technology Review. He completed a doctorate at MIT in cognitive psychology and received a bachelor’s degree in symbolic systems from Stanford.