摘要: We describe a direct analog implementation of a neural network model of olfactoryprocessing. This model has been shown capable of performing hierarchical clustering as a result of a coactivity-based unsupervised learning rule whic... 展开 We describe a direct analog implementation of a neural network model of olfactoryprocessing. This model has been shown capable of performing hierarchical clustering as a result of a coactivity-based unsupervised learning rule which is modeled after long-term synaptic potentiation. Network function is statistically based and does not require highly precise weights or other components. We present current-mode circuit designs to implement the required functions in CMOS integrated circuitry, and propose the use of floating-gate MOS transistors for modifiable, nonvolatile interconnections weights. Methods for arrangement of these weights into a sparse pseudorandom interconnection matrix, and for parallel implementation of the learning rule, are described. Test results from functional blocks on first silicon are presented. It is estimated that a network with upwards of 50K weights and with submicrosecond settling times could be built with a conventional CMOS double-poly process and die size Olfactory, Synchronous analog, Granger/lynch. 收起