A Chaotic Feature Extracting BAM and Its Application in Implementing Memory Search

作者：Nejadgholi Isar; Seyyedsalehi Seyyed Ali^*; Chartier Sylvain

来源：Neural Processing Letters, 2012, 36(1): 69-99.

DOI：10.1007/s11063-012-9223-3

• Neural network
• Chaos theory
• Output variability
• Cognitive science
• Associative memory

An unsupervised trained, chaotic BAM composed of units with saturated limits logistic function can extract intrinsic features of data and exhibit various associative dynamics for different values of transmission parameters of the neurons%26apos; activation functions during recall. The output behavior of its units can be a fixed point or periodic attractor, a constrained aperiodic attractor consisting of one or more stored patterns, or a chaotic attractor. This characteristic indicates that the model is a promising technique that can be applied to information processing, such as pattern recognition and memory recall. However, controlling the amount of output variability and stabilizing it in a desired attractor is a crucial issue in practice. In this work it is shown that the transmission parameters of the units%26apos; activation functions play a significant role in identifying the output behavior. Using different time-series generated by the trained network, Largest Lyapunov Exponent is computed for different values of transmission parameter. Then, critical values of this parameter that lead to the highest chaotic behavior for each unit are stored and used to set the network during recall. Interaction between some chaotic feature units and some fixed-point ones produces desired behaviors with various degrees of uncertainty. An evolutionary algorithm is then introduced to find the units that should work chaotically to generate the desired behavior. Achievement of this method implies that such a controlled chaotic feature extracting BAM can be feasibly applied to information processing.

• 出版日期2012-8