Figure 4.

Coarse-Coded Conjunctive scheme results in similar yet distinct representations for related input stimuli. This allows the culture to do pattern generalization and novel pattern classification. Such a scheme is also shown to be robust to probabilistic nature of connections. (a) Output Classification accuracy for the 56 paired stimulus patterns over different trials in different cultures (C) studied on different days in culture (DIV, days in vitro). The numbers indicate the number of patterns classified with an accuracy greater than 80%. The results are shown separately for a time delay of 3ms and 0.5ms of pairing. (b) Classification ability of the cultures for a novel pattern. 8 patterns, one from each group of paired stimuli (indicated for example in Fig. 2 were systematically left out and the perceptrons were trained to classify rest of the inputs into different group. The height of the bar indicates the average number of different such hidden patterns (out of 8) that were correctly classified (with greater that 80% accuracy (Chance = 1/8)). This indicates that patterns are grouped into linearly separable groups in higher dimensions based on the first electrode stimulated. (c) The coding of the outputs is such that a perceptron connected probabilistically to a fraction of output electrodes is able to classify the inputs without significant degradation. The curve is averaged over 11 trials in Fig. 4a. It indicates the reduction in number of input classes correctly classified with greater that 80% accuracy as the number of connections each perceptron receives is reduced. The blue trace indicates the loss of accuracy when perceptrons are connected randomly to a fraction of output electrodes (randomized 3 times and mean number of classes calculated). This loss of accuracy can be seen as illustrated in Fig. 1(f–h).