Adaptive Digital Learning Network

This classification procedure was originally developed by Bledsoe and Browning C3721 and first applied to chemical problems by Stonham/ Aleksander, et.al. C2843. The learning network has some similarities to the perceptron, especially in the random combination of features 066, 3673. 

The patterns used with this method are binary encoded patterns x. Groups Cn-tuples) of n (usually 3 or 4) pattern components are randomly chosen and associated with a memory element . A memory element has 2 addressable 1-bit-storage locations. The configuration of the bit pattern of the n-tuple is used to address one of the 2 locations. Example For n = 4 a four-bit binary pattern is interpreted as one of the decimal numbers 0 to 15. In the training stage a 1 is written into the location addressed by the n-tuple. 

A digital learning network consists of a group of memory elements. For example, if patterns with 100 features are to be classified and n = 4, then a set of 25 16-bit-memory elements are necessary (with no feature being sampled twice). All elements are initialized to 0 and connected randomly to the pattern components. 

To train a digital learning network for a particular class, only patterns belonging to that class are presented to the network. Each n-tuple of features selects a storage location. If the selected storage already contains a 1 nothing happens, otherwise a 1 is written (an alternative training method is described below). 

The digital learning network may be implemented with hardware or simulated with a computer program. Several characteristics of the method have been investigated by Stonham et.al. C282 - 285H. 

---

Stonham TJ, Aleksander I, Camp M, Pike WT, Shaw MA (1975) Classification of mass spectra using adaptive digital learning networks. Anal Chem 47 1817... 

FIGURE 36. Generation of a reduced optimum training set for the adaptive digital Learning network E2851. 

Early optimism by Stonham, Aleksander, et. al. C282, 283, 284, 2853 about the usefulness of adaptive digital learning networks could not be confirmed by Wilkins et. al. C2803. 

---