Network Operation

The heart of the ART networks is a scheme for interaction between the units in the output layer which guarantees that the most active unit will suppress the activity of all the other units. After the output units have been activated, they are allowed to interact until only one remains active. This unit then activates the units in the first layer via another set of paths with their own weights. The input vector is then compared with the pattern that has been passed down from the output layer.

The comparison between these patterns is controlled by a parameter called the vigilance threshold, which determines the degree of similarity between patterns that are given the same classification. If the difference between the input pattern and the pattern that is passed down from the output layer is less than the vigilance threshold, then the active output unit is accepted as determining the correct classification of the input, and the weights associated with that unit are adjusted according to one of the schemes described below.

If the difference between the input pattern and the pattern passed down from the output layer is greater than the vigilance threshold, the activity of the active unit is completely suppressed, and process described above is repeated until either a classification is achieved, or there are no units left, in which case a new class is created by the addition of a unit to the output layer.

The weights that connect each output unit to the input units represent the pattern typical of the class to which the output unit belongs. The weights that connect the input units to each output unit represent the same pattern, except that their values are normalized.

The effect of the operation is as follows: for each input vector, the network finds the closest pattern among the output units. If this is within the distance determined by the vigilance threshold, this is accepted as the correct classification and the weights adjusted to make the stored pattern more similar to the input vector. Otherwise, the network continues to go through the remaining output units in order of similarity to the input vector. If any of them is found to be within the distance determined by the vigilance threshold, it is chosen as the correct classification. If no output pattern is sufficiently similar, a new output unit is created.

Both ART1 and ART2 operate in this way. The only difference between the two is that ART1 inputs consist of binary patterns, while the inputs to ART2 networks can have arbitrary values, which are normalized by additional operations that are applied within the input units.