Adaptive Switching Circuits

Abstract

Ada^livft or "ieuming" ayatemB can automatlc-illy modify their ovm structures to OY^imizt performance based on past experiences. The system desiwier "teaches" by shoving the system examples of l..puc slyiilo or pattemp and simultaneously what he would like the output to be for each input. The cyBtem In turn organizes Itself to comply as wall as possible with the wishes of the designer. An adaptive pattern classification machine (called "Adallne", for adaptive linear) hac been devised to illustrate adaptive behavior and artificial learning. During a training phase, crude geometric patteri.3 arp fed to the machine by setting the toggle swJt.hes in a Vx 1 * input array. Setting another toss^e switch tells the machine whether the desired output for the particular input pattern is +1 or-1. All input patterns are classified into two categories. The system learns a little from each pattern and accordingly experiences a design change. After training, the machine can be used to classify the original patterns and noisy (distorted) versionn of these patterns.

A competitive neural model of small number detection

“…Whilst the SOM uses unsupervised learning, the mapping of the number line to a precise value is achieved using a single layer perceptron (SLP), trained using the delta learning rule (Widrow & Hoff, 1960). This mapping models the association of precise values to a foundation of numerical magnitudes, with such transcoding central to all of the models of numerical processing discussed.…”

Retrospective Revaluation as Simple Associative Learning.

“…Second, to determine the weights corresponding to a given learning experiment we need a learning algorithm. Learning is modeled here with a refinement of the RW learning rule, first derived by Widrow and Hoff (1960) and introduced to animal learning theory by Blough (1975). In this algorithm weight W i changes according to…”

Adaptive, integrated sensor processing to compensate for drift and uncertainty: a stochastic ‘neural’ approach

“…They are the weights w ij , the noise control parameters, a j of the CRBM and the weights w k in the output SLP layer. The CRBM parameters are optimised by minimising contrastive divergence (MCD) [11] while the SLP is trained using the delta rule [16].…”