Neural networks with color neurons and hidden units: memory without errors and attention ability

The author proposes and studies a non-Hopfield model which makes it possible to apply theoretical physics tools. It is a two-layer neural network containing two types of neurons. The first type is described by continuous scalar functions. The second type is described by multicomponent vector functions. Since for recognition of color pictures the different components of the state vectors might correspond to the different components of the color spectrum, one can call them color neurons. Such neural networks can discern correlated patterns, admit local learning rules, have large enough memory without spurious states, and possess other useful properties, including the cognitive ability to distinguish whether or not an input pattern is far away from any of the embedded patterns and certain elements of attention