The brain is a remarkable piece of work. At a given moment, from a blizzard of incoming data - visual, tactile, auditory, olfactory, taste, memory, etc. - it knows instantly how to classify what information it wants, and discard or store the rest. One sound in a roomful of noise. One object in complex scene. One expression in a hundred faces. The brain seems to be able to do this in a non-linear fashion. If we could get our sensors to do it just as well and just as fast, then problems might be solved that today defeat even the fastest computers.
This was the starting point for research at the Salk Institute funded by the Office of Naval Research. How can you separate signals that have been mixed together, without knowing ahead of time what those signals might be? The general approach to this problem, called Independent Component Analysis, had been pioneered in France in the 1990s, but it was Tony Bell, working with Terry Sejnowski in the Computational Neurobiology laboratory at Salk who came up with a fast, practical algorithm, called the infomax ICA.
"It's all about the signal to noise ratio, and whether this kind of data can be analyzed and processed using traditional methods," says Dr. Joel Davis, manager of the project in ONR's Cognitive, Neural, and Biomolecular S & T Division. Can ICA be used to eavesdrop on the brain itself? In an example of serendipity, Scott Makeig (at the Naval Health Research Center) working with Bell and Sejnowski, discovered that the ICA algorithm is extremely effective at separating the sources of weak and noisy independent electrical signals from deep within brain that are measured on the scalp (EEGs). Previous methods for analyzing EEGs had relied on averaging over many trials, but with ICA it was possible to analyze single trials.
"Pattern recognition and extracting signals from noise has always been difficult, especially under high noise conditions," says Davis. "ICA is changing the whole field of event-related potentials, and this is not trivial. In a room full of bells, whistles, music, clapping, and human voices, traditional signal processing may fail to recognize the human voice. Not so ICA. The signals and information our brains care about - like the human voice - are non-linear, as is the language of the brain itself. Using the ICA algorithm really works," Davis said.
The patented algorithm is now being applied to dozens of applications. It is currently available on the Web for any who wants to explore its use for science and research purposes at: http://www.cnl.salk.edu/~tewon/ica_cnl.html. Users are asked to note their applications. To date, ICA has been applied to over a hundred real world problems including speech and pattern recognition, telecommunications, satellite imagery, sound spectrograms, image processing, de-noising, ultrasound research, spacecraft fault detection, astronomy, brain-computer interfaces, and analyses of climactic data.