How does the brain register a sensory clue, such as the sound of footsteps approaching in a dark alley, and signal us to take necessary action? After all, it is our ability to make such decisions — to rapidly avoid trouble or to secure a reward — that ensures our survival.
Robert Gütig | Jan Ficner
In a report published in the journal Science last year, Robert Gütig, a research leader in theoretical neuroscience at the Max Planck Institute of Experimental Medicine in Göttingen, Germany, put forward a new model explaining this phenomenon, earning the 2016 AAAS Newcomb Cleveland Prize.
The Association's oldest prize, now supported by The Fodor Family Trust, annually recognizes the author(s) of an outstanding paper published in the Research Articles or Reports sections of the journal Science.
Gütig's prize-winning model, called aggregate-label learning, represents "a major advance in machine learning and systems neuroscience," said Peter Stern, the paper's editor. "The question of how an organism can associate a sensory signal with a crucial behavioral outcome or action that may become immediately necessary has been a critical one."
Indeed, for years, scientists have sought to better understand how humans and animals bridge the gap between hearing or observing a sensory clue that signifies a coming event, like the sound of a twig breaking in the forest as a bear approaches, and taking action, such as fleeing from the bear.
"I think difficulties in this space can be traced back to two long-standing questions," Gütig explained. "Firstly, scientists have not fully understood how much information the teaching signals that drive learning in the brain provide about given sensory clues. And secondly, how a neuron adjusts its synaptic efficacies to most efficiently change response to a certain pattern of input activity has been unknown."
In general, a neuron spikes, meaning it it sends an electrical signal that travels down a nerve fiber, when cued by a sensory element. Neurons that respond to random inputs typically generate spike outputs that are not in specific patterns and thus do not convey a specific action to the person or animal.
In his model, Gütig proposes that neurons can serve as reliable clue detectors to their host by matching the number of spikes they generate to the magnitude of a clue they perceive. The sum of the clues conveyed creates a neural pattern that signifies a danger or opportunity.
In a series of "learning trials," Gütig compared his model to others that have been proposed to describe neural learning behavior.
Gütig's model of aggregate-label learning could be used to streamline technologies such as automatic speech recognition by phones or processing of simple text to create subtitles for TV shows.
In these tests, Gütig explained, "a neuron received a continuous stream of sensory clues represented by different input patterns. In the real world, these sensory clues could be a stream of words or a sequence of images. For some clues, the neuron is expected to respond and for others it is expected to remain silent. The number of learning trials I did measured how often the neuron needed to be trained until it reached 90% correct responses."
Gütig found that his model outperformed the counterpart models, requiring fewer neuron signaling connections and learning trials to figure out what was going to happen next.
His model could have a variety of uses. "For technological applications, it promises a substantial simplification of the training data required for automated processing of continuous streams of speech," he said. Such processes include automatic speech recognition by phones, for example, or processing of simple text to create subtitles for TV shows.
"Traditionally," he explained, "if you want to train an automatic speech recognition system to detect a word like 'Siri,' you need to have a training data set for which you know where in time this word appears … In my model, you only need to know how often the word appeared, but not when. This information is much easier and cheaper to obtain … In essence, my approach shows that word detectors can be trained on the basis of word counts only."
Gütig said he was inspired to do research in this space, because like many scientists, he wanted to better comprehend how the brain works. "In my case, I am particularly fascinated by understanding what algorithm is running inside our brains," he said. "I would love to have a print-out of its source code. I therefore study the computational principles that underlie information processing and learning in spiking neural networks."
The Newcomb Cleveland Prize was established in 1923 with funds donated by Newcomb Cleveland of New York City and was originally called the AAAS Thousand Dollar Prize. Along with a plaque and $25,000 in prize money, the winner receives complimentary registration and reimbursed travel expenses to attend the AAAS Annual Meeting.
The 2015-2016 Newcomb Cleveland Prize Selection Committee includes Jeremy Berg, Chair, editor-in-chief, Science; Gyorgy Buzsaki, New York University School of Medicine; Jennifer Doudna, University of California, Berkeley; Harinder Singh, Cincinnati Children's Hospital Medical Center; and Maria Zuber, Massachusetts Institute of Technology.
The Newcomb Cleveland Prize will be presented to Robert Gütig at the 183rd AAAS Annual Meeting in Boston, Massachusetts, which will take place 16-20 February. The AAAS Awards Ceremony and Reception will be held at 6:30 p.m. on 17 February in the Sheraton Boston Hotel.
[Credit for associated image: Chris Bickel/ Science]