As computer power has increased, the large number of labeled images on the Internet has allowed computer programs to train their “neural networks” to identify objects in photographs, recognize speech in cell phones, help self-driving cars navigate and even beat humans in video games.
But the basic architectural design and computer algorithms of current neural networks have remained the same since the 1980s. They use a one-way layer-by-layer process to analyze information and learn how to identify images.
Now with a $12 million grant over five years, Carnegie Mellon University will study the mouse brain to better understand complex neural networks and feedback loops it, and our own brains, use in processing visual information. Once understood, those details could lead to mathematical models and algorithms to advance machine learning and artificial intelligence.
The project is part of President Barack Obama’s brain initiative programs to revolutionize understanding of the brain. The federal Intelligence Advanced Research Projects Activity, through its Machine Intelligence from Cortical Networks program, or MiCrOns, awarded the grant that seeks to achieve “a quantum leap in machine learning” through “novel computer programs inspired by brain architecture.”
The CMU-based team led by Tai Sing Lee, a professor in the computer science department and the Center for the Neural Basis of Cognition, has been working with colleagues to develop theories on how the brain processes visual information.
In current methods of machine learning, teaching a neural network to identify a car or a face involved the input of thousands of images along with the correct image. Any time the computer says it’s something other than the target image, the error gets funneled back through the network to make changes in the connections so that the network gradually will learn to produce the correct answer. That process is known as supervised learning.
But humans and animals can learn without supervision, and often with only a few examples. Scientists believe that the secret to this ability lies in the enormous number of feedback loops between and within each stage of information processing in the brain’s neural networks.
Mr. Lee, who holds a doctorate in engineering science and medical physics, said once our brain receives visual information from the retina, it detects simple features and then successively more complex features to interpret what our eyes might be seeing. The feedback loops can be used to imagine what the future incoming data might be. When the imagination and the prediction match the incoming sensory data, the brain knows it understands the scene correctly.
“In a way, the brain might function like a detective or a scientist,” Mr. Lee said. “It makes observations and formulates some theories and models, and then imagines the scenarios that could give rise to the observed evidence. Furthermore, once it has an interpretation based on current evidence, it will look for further evidence to get a clearer picture.”
For example, you might see what’s initially thought to be a face. Further analysis identifies wrinkles and eyes, among other details, eventually providing enough detail to verify not only a face but also a specific person. “Understanding the neural circuits underlying these processes could allow us to empower machine-learning and computer-vision systems to see more like humans do with imagination,” he said.
Mr. Lee will work with Sandra Kuhlman, a CMU assistant professor of biological sciences, to record signals from many thousands of neurons in areas of the mouse brain that process visual information. In the past, only a few neurons could be sampled during such experiments.
Next, a Harvard University team will reconstruct the complex circuitry of neurons that the CMU scientists record. Finally, CMU computer scientists will study how those neurons are connected and how they behave, then test before selecting the best theories on how the brain processes information.
The MiCrOns project “is the most amazing part of the Obama Brain Initiative,” said George Church, professor of genetics and health sciences and technology at Harvard University Medical School. “There are three teams focused on the same cubic millimeter of visual cortex in the brain,” he said, noting CMU’s pioneering efforts in the project.
Mr. Lee said there’s hope this information will reveal the principles and the mechanisms the brain uses to learn and to process information with feedback loops. “Understanding these principles will allow us to program computers to learn like a baby, without much supervision and often with a few examples, and more importantly, to the see the world with imagination.”
David Templeton: dtempleton@post-gazette.com or 412-263-1578.
First Published: February 23, 2016, 5:00 a.m.
