Human Brain Model Recognizes, Writes, and Remembers
A new model of the brain with 2.5 million simulated neurons bridges the gap between complex mental activity and sophisticated behaviors.
-- by Heather Kathryn Ross
“Spawn” brings to mind demonic comic book heroes and Hollywood super-viruses, but Spaun—a computer model of the human brain with 2.5 million working “neurons”—is more akin to Steven Spielberg’s A.I.
Spaun, which stands for Semantic Pointer Architecture Unified Network, is shedding light on an important neuroscientific quandary: how activity in the brain actually translates into intricate, measurable behaviors.
Researchers in Ontario, Canada presented Spaun with images of typed or handwritten numbers, and the model brain used the data to perform a series of eight tasks with its robotic arm. Spaun could recognize and replicate numbers, calculate a potential reward during a “gambling” exercise, remember a list of items, and create and complete patterns, demonstrating working memory, reasoning, and reinforcement learning.
Spaun performed as well as a human brain at recognizing patterns between two groups of numbers and constructing a corresponding third group—no simple feat.
“Everyone knows the brain is complex. Everyone knows that behavior is also complex. No one knows how the complex activity in the brain gives rise to that complex behavior,” said Dr. Chris Eliasmith, lead researcher and director of the Centre for Theoretical Neuroscience at the University of Waterloo. “Spaun is the first large-scale brain model that is both complex in its own right and gives rise to sophisticated behavior.”
The Expert Take
Spaun is intriguing not just for its sophistication, but also because, by incorporating a wealth of existing knowledge about the structure and function of specific regions of the human brain, it is as close to artificial intelligence as researchers have yet come.
“By assembling a large amount of brain know-how into one model, Eliasmith [and colleagues] have provided a coherent theory of how the brain works,” wrote Christian Machens of the Champalimaud Centre for the Unknown in Portugal in an article that will run alongside Eliasmith’s findings in the latest issue of Science. “Spaun levels the playing field by setting a new goal and a new benchmark for [brain] simulations: to not simply incorporate the largest number of neurons or the greatest amount of detail, but to reproduce the largest amount of functionality and behavior.”
Though groundbreaking, Spaun falls far short of a real, adaptive human mind. For instance, the inputs it is able to recognize and process are limited to the digits 0 through 9. The model also does not include all areas of the human brain, and those that are integrated perform only a few of the thousands of functions our brains execute every day.
“The Spaun model is much simpler than a human brain—it only has 2.5 million neurons, compared to 100 billion in the human brain,” Eliasmith said. “Spaun does have several kinds of learning in it (for example, in the reinforcement learning task, and in the pattern completion tasks). However, Spaun cannot learn a completely new task from scratch. This is one of the areas of focus for our future research.”
Source and Method
Sections of the Spaun computational model were constructed to mimic the physiological structures of specific areas of the human brain, such as the primary motor cortex.
In order to perform its eight tasks, Spaun was presented with hundreds of images of handwritten or typed characters. It evaluated the characters using an “eye” similar to a camera lens, and used a physically modeled arm to perform tasks of varying complexity, from image recognition and copying to fluid reasoning.
Because Spaun’s neural architecture is similar to that of a real human brain, its success in performing complicated tasks proves that computer models of the mind can be flexible within the constraints of human biology.
“Spaun has suggested that the particular architecture and kind of neural representations used in the model are quite plausible,” Eliasmith said. “This is because it matches many sources of data about brains, including single cell activity, high-level behavior, and neuroanatomy and physiology.”
The Spaun model offers exciting possibilities for medical researchers exploring how changes in specific neurons in the brain affect behavior. Eliasmith has invited other researchers to further his discoveries by making Spaun's specs public.
“We have recently submitted work on the effects of neuron loss during aging to cognitive decline. This provides a clear example of how we can use models like Spaun to better understand the link between changes in the brain and changes in behavior,” Eliasmith explained. “The model itself can be downloaded from our website, and other researchers are free to use it and the software that we used to build this model. The implications of various kinds of damage or chemical changes to the model can be explored by simulating it under different conditions.”
Spaun is not the first or the largest neural simulation model, but it is the first devoted to “bridging the brain-behavior gap.”
In 2005, the Blue Brain Project at École Polytechnique Fédérale de Lausanne in Switzerland was the first large-scale attempt to recreate a mammal’s brain using a supercomputer. Incorporating more than one million simulated neurons in cortical columns, the Blue Brain team has made significant strides toward creating a complete virtual model of the human mind.
Since 2007, a team at IBM’s Almaden Research Center has been working on a cognitive computation project to recreate a 55 million- to- one billion-neuron cat-scale brain to offer even greater insights into the inner workings of mammalian intellect.