Georgia Tech inventors have created a compact and programmable analog that is a multi-dimensional radial basis function (RBF) classifier. Each component of the feature is modeled by a Gaussian distribution, which allows the mean and variance of the invention to be independently programmable and stored inside a floating-gate bump cell. The performance of the RBF based circuit is comparable to that of a digital counterpart and the efficiency of this approach can be at least two orders of magnitude better.
- Electric devices
Although conventional radial basis function (RBF) classifiers work in certain limited applications, they are significantly inadequate in fully approximating the Gaussian function because the conventional classifiers cannot adjust the width of the transfer curve. Another drawback of conventional devices is that they require extra hardware to store or to periodically refresh template data when they are employed in a recognition system. Therefore, conventional analog classifiers are inaccurate and fail to provide statistic information that significantly reduces the amount of digital processing required.