No one knows how to directly ascertain the brain’s learning rules, but there are many highly suggestive similarities between the brain’s behavior and that of the Boltzmann machine.
Both learn with no supervision other than the patterns that naturally exist in data. “You don’t get millions of examples of your mother telling you what’s in an image,” Hinton said. “You have to learn to recognize things without anybody telling you what the things are. Then after you learn the categories, people tell you the names of these categories. So kids learn about dogs and cats and then they learn that dogs are called ‘dogs’ and cats are called ‘cats.’ ”
Adult brains are less malleable than juvenile ones, much as a Boltzmann machine trained with 100,000 car images won’t change much upon seeing another: Its synapses already have the correct weights to categorize a car. And yet, learning never ends. New information can still be integrated into the structure of both brains and Boltzmann machines.
Over the past five to 10 years, studies of brain activity during sleep have provided some of the first direct evidence that the brain employs a Boltzmann-like learning algorithm in order to integrate new information and memories into its structure. Neuroscientists have long known that sleep plays an important role in memory consolidation, helping to integrate newly learned information. In 1995, Hinton and colleagues proposed that sleep serves the same function as the baseline component of the algorithm, the rate of neural activity in the absence of input.
“What you’re doing during sleep is you’re just figuring out the base rate,” Hinton said. “You’re figuring out how correlated would these neurons be if the system were running by itself. And then if the neurons are more correlated than that, increase the weight between them. And if they’re less correlated than that, decrease the weight between them.”
At the level of the synapses, “this algorithm can be implemented in several different ways,” said Sejnowski, who earlier this year became an advisor on the Obama administration’s new BRAIN Initiative, a $100 million research effort to develop new techniques for studying the brain.
The easiest way for the brain to run the Boltzmann algorithm, he said, is to switch from beefing synapses up during the day to whittling them down during the night. Giulio Tononi, head of the Center for Sleep and Consciousness at the University of Wisconsin-Madison, has found that gene expression inside synapses changes in a way that supports this hypothesis: Genes involved in synaptic growth are more active during the day, and those involved in synaptic pruning are more active during sleep.
Alternatively, “the baseline could be computed during sleep and changes made relative to it during the day,” Sejnowski said. His lab is building detailed computer models of synapses and the networks they sustain in order to determine how they collect firing statistics during wakefulness and sleep and when they change synaptic strengths to reflect the difference.
A Boltzmann-like algorithm may be only one of many that the brain employs to tweak its synapses. In the 1990s, several independent groups developed a theoretical model of how the visual system efficiently encodes the flood of information striking the retina. The theory held that a process similar to image compression called “sparse coding” took place in the lowest layers of the visual cortex, making later stages of the visual system more efficient.