Abstract: Network algorithms and architectures for patternrecognition derived from neural models of the olfactorysystem are reviewed. These span a range from highlyabstract to physiologically detailed, and employ thekind of dynamical complexity observed in olfactorycortex, ranging from oscillation to chaos. A simplearchitecture and algorithm for analytically guaranteedassociative memory storage of analog patterns,continuous sequences, and chaotic attractors in thesame network is described. A matrix inversiondetermines network weights, given prototype patterns tobe stored. There are N units of capacity in an N nodenetwork with 3N$+2$/ weights. It costs one unit perstatic attractor, two per Fourier component of eachsequence, and three to four per chaotic attractor.There are no spurious attractors, and for sequencesthere is a Liapunov function in a special coordinatesystem which governs the approach of transient statesto stored trajectories. Unsupervised or supervisedincremental learning algorithms for patternclassification, such as competitive learning orbootstrap Widrow-Hoff can easily be implemented. Thearchitecture can be 'folded' into a recurrent networkwith higher order weights that can be used as a modelof cortex that stores oscillatory and chaoticattractors by a Hebb rule. Network performance isdemonstrated by application to the problem of real-timehandwritten digit recognition. An effective system withon-line learning has been written by Eeckman and Bairdfor the Macintosh. It utilizes static, oscillatory,and/or chaotic attractors of two kinds -Lorenzeattractors, or attractors resulting from chaoticallyinteracting oscillatory modes. The successfulapplication to an industrial pattern recognitionproblem of a network architecture of considerablephysiological and dynamical complexity, developed byFreeman and Yao, is described. The data sets of theproblem come in three classes of difficulty, andperformance of the biological network is favorablycompared with that of several other network andstatistical pattern recognition methods.!
展开▼