One of the fundamental goals of cognitive and systems neuroscience is to create a computer program that can simulate the activity of the human brain, from single neurons, through network level processing, to influences on behavior. The only problem is that the human brain contains almost 90 billion neurons with an estimated total of 100 trillion synapses! As staggeringly large as those numbers are, researchers actually aren’t too far away from producing such a complex computer program.
Recently, Canadian researchers created a computer model that can see images, recognize patterns, and write out answers with a robotic arm. Their model, called the Semantic Pointer Architecture Unified Network (or “Spaun”), can perform 8 different tasks, including activities in perception (image recognition), pattern recognition (counting), working memory (memorizing numbers), and reinforcement learning (answering IQ test questions). And if that weren’t impressive enough, Spaun performs all these tasks with only 2.5 million simulated neurons!
In the past decade, a large amount of neuroscience research has focused on using computers to model the activity of the brain. For the most part, those researchers have attempted to expand the number of cells the models include while maintaining the biological accuracy of the simulations. What makes Spaun different from all the previous computer models is that it produces complex behaviors with fewer neurons. Moreover, Spaun simulates the electrical and chemical physiology of each neuron, and organizes each neuron into groups to represent specific brain regions that are involved in visual perception, decision-making, short-term memory, and movement.
Spaun is made up of two main components: a working memory system, which mimics the activity of the prefrontal cortex, and an action selection system, which mimics the activity of the basal ganglia and thalamus. Using an amalgam of neural integrators taken from computation neuroscience, Spaun can store information as well as bind newly arriving information with already stored information. In this way, Spaun is actually quite human, remembering the first and last items on a list better than any others!
The model is not perfect, though. While Spaun can perform those set of tasks with ease, it cannot learn a new one. Also, the computation is much slower than our brain’s processing, and some of the simulated neurons have unrealistic spiking statistics. Perhaps these shortcomings will be remedied when we have a better understanding of the brain. Instead of using a bottom-up approach to model the brain by replicating its parts, connections, and organization, future large-scale neural simulations will most likely use a top-down approach to capture the brain’s complex behaviors. Perhaps a next model will be able to acquire new information based on experience, just like humans do! Until then, the race is on to create a computer simulation of all the neurons of the brain that can reproduce the greatest amount of functionality and behavior!Eliasmith C., Stewart T.C., Choo X., Bekolay T., DeWolf T., Tang Y. & Rasmussen D. (2012). A Large-Scale Model of the Functioning Brain, Science, 338 (6111) 1202-1205. DOI: 10.1126/science.1225266 Image adapted from Andrew Brookes/Corbis and Andrzej Wojcicki/Science Photo Library/Corbis.