ALBUQUERQUE, N.M. The human brain nearly instantaneously classifies objects, even when parts of them are hidden by obstructions. Computer-based pattern recognition, on the other hand, rarely operates in real-time. But Sandia National Laboratories researchers believe they have found a way to level the playing field.
In hundreds of human trials, the researchers empirically measured the visual template that they believe the human brain uses for associating parts of an object. Based on those measurements, they claim to have statistically verified a template for real-time pattern recognition that could be used by computer vision systems.
"Since our classification method is based on human perception rather than mathematical equations, it is almost too simple to explain to those of us who expect complexity," said physicist Gordon Osbourn.
Osbourn's department at Sandia was researching quick, errorless identification methods for its handheld "lab-on-a-chip" chemical-sensor system when it derived the technique. Sandia researchers John Bartholomew, Tony Ricco and Greg Frye applied the method to multidimensional data coming in from the lab-on-a-chip chemical sniffer and recently detailed their successes in a scholarly paper.
One unexpected advantage the researchers found was that data need not come in nice, statistically significant batches for the technique to work. Indeed, the researchers reported, data can come in fits and starts or be clumped together in nearly any type of distribution.
"Conventional pattern-recognition algorithms often require that your data be collected in Gaussian distributions to work well, but modern sensor arrays can often relay data in non-uniform distributions, causing false alarms," Osbourn remarked.
Sandia's lab-on-a-chip is aimed at detecting chemical weapons on the battlefield so that soldiers have sufficient warning to don protective equipment. It could also be used to ferret out hidden explosives in airports as luggage passes through automatic handlers. "There are many other applications that we haven't developed yet. For instance, physicians often need to quickly analyze complicated medical images, and certain environmental analyses also require fast, accurate results," said Osbourn.
Today's pattern-classification algorithms can be too complex for real-time operation, but biological brains, performing fewer operations per second than silicon chips, can accomplish such tasks almost instantaneously. Since the recognition tasks occur without any conscious thought, Sandia Labs researchers reasoned that a low-level template-matching operation must be performed automatically.
To measure the shape of that hypothetical low-level template,the group performed hundreds of human trials during which people quickly grouped real-world objects seen near each other.
After extensive testing, Osbourn's group surmised that people superimpose a dumbbell-shaped pattern over any two points to determine whether the points belong to the same object or to different objects. If both points fit inside the dumbbell shape without having to include other extraneous points, then they belong to the same object. But if extraneous points have to be included, the brain concludes that the two points do not belong to the same object.
"This is happening on a subconscious level, which is why it's so fast. The subconscious mind sizes the dumbbell so that each end is centered on one point. If no other point intrudes in that space, then one considers the two points a group," said Osbourn.
Though the group derived its simple template-matching mechanism from two-dimensional visual experiments, the researchers have discovered that it can be applied to any sort of sensor data and in any number of dimensions. In effect, the mechanism permits computers to "see" sounds, smells or even combinations of sensor inputs forming multidimensional spaces. The computer merely applies the same empirical judgments based on the proximity of points in a data set, regardless of what the points represent or in how many dimensions they exist.
"The mathematical transformation from two-dimensional vision data to multidimensional data from different types of sensors is very straightforward and seems to work just as well as it did for vision," said Osbourn.
Sandia Labs has applied for a patent covering the use of the algorithm, which the researchers call Veri (visual empirical region of influence). To help users apply Veri, Sandia researchers have written software that symbolically represents data points as dots, which are connected to outline objects. Researchers familiar with the technique claim that relationships among previously disparate data sets are immediately illuminated by the connecting-the-dots method.
To confirm Veri's ability to recognize real-world distributions, Osbourn's group tested it on 25 patterns from the computer-science literature that have traditionally proved troublesome for various pattern-clustering algorithms. Osbourn claimed Veri outperformed all commercial clustering algorithms and resulted in the fewest false alarms.
"Instead of just classifying an unknown sensor input in the closest matching category, as neural nets can sometimes do when there is a lot of noise in the data, Veri will recognize an unknown input as something new. In this way, Veri minimizes the number of false alarms it causes," said Osbourn.