Researchers Develop Algorithms to Make Sense of and Mimic the Brain
It's 3:30 p.m. on a weekday afternoon and members of the Salk's Computational Neurobiology Laboratory (CNL) are starting to migrate toward their daily meeting place.
The area has a freshwater fish tank, two long sofas arranged in an L and a round table spread with crackers, chocolate covered almonds and hot tea for everyone.
It's a place where they converge to relax and talk about what's on their minds—a conference, an interesting article, their experiments, and in some cases, their plans to turn research into products that will benefit the public.
Recently, these conversations were about SoftMax, a company acquired in December by wireless communications giant Qualcomm for its proprietary noise reduction technology. The algorithm behind the technology was developed in the CNL and is used to suppress background noise and make phone calls more intelligible.
But what does a basic biological research lab at Salk have to do with cell phones? Terry Sejnowski, professor and head of the CNL, explains.
"Our research is primarily focused on learning about the human brain and what it can do, and it's about learning how to make computers that are able to mimic these processes. It's a two-way street."
The noise-reduction algorithm that led to SoftMax is based on independent component analysis (ICA), a type of mathematical formula developed by Tony Bell, a former postdoctoral fellow in Sejnowski's lab. The algorithm solved the cocktail party problem: How to separate out a single sound source from mixtures of recorded signals.
SoftMax independently developed a version of ICA that mimics the human auditory system. It uses a pair of microphones in a cell phone headset, much like a pair of ears, to measure differences in loudness and timing to separate the source of a voice from the background.
"It identifies a speaker's voice from every other sound, and simply intensifies the voice signal while dampening the others," explains Sejnowski, who was on SoftMax's scientific board.
Alternative ICA algorithms for sound reduction exist, he notes, but the SoftMax version is more robust because it can be used with two, 10, or 100 microphones for better sound resolution.
ICA can be used with any type of signal, not just sound. One day, for example, it may help machines read electrocardiograms to reduce the time cardiologists spend with paper recordings and computers. It's already being used to understand what segments of our DNA are being transcribed, and what simply represents interfering background noise.
Philip Low is a postdoctoral fellow in the CNL who has been using electroencephalograms (EEG) to study sleep in birds and humans. With Sleep Parametric EEG Automated Recognition System (SPEARS), an algorithm he created at CNL during his doctorate studies, Low has been able to study brain activity using only two electrodes placed on the scalp, not the dozens that are normally placed all over the head and body in sleep studies.
Furthermore, using SPEARS, he has identified a new range of electrical activity, or "sleep state," that was previously masked by the low frequency waves produced during sleep. This new technology has led Drs. Low and Sejnowski to form NeuroVigil, a company that revolutionizes the way brain electrical activity is both recorded and analyzed.
Through a collaborative consortium with major academic institutions, NeuroVigil plans to perform basic research for the iBrain, an iPod for the brain which will monitor people's health in real-time.
"I've been very fortunate because the members of my lab are exceptionally talented," says Sejnowski. "We form an eclectic team, with backgrounds in mathematics, physics, electrical engineering, psychology, and even philosophy. But this is what's required to even approach solving the big problems of how the brain works."