Adaptive Resonance Theory/ART Adaptive Resonance Theory

ART Definition

ART is a theoretical model that can actively generate and organize environmental information and cognitive coding in neural networks when the neural network interacts with the environment.

ART Advantages

1. It can conduct real-time learning and adapt to non-stationary environments.

2. It has a stable and rapid recognition ability for objects that have already been learned; at the same time, it can also quickly adapt to new objects that have not been learned.

3. It has the ability of self-normalization. Depending on the proportion of certain features in the whole, some features are sometimes treated as key features and sometimes as noise.

4. It does not need to know the sample results in advance, which is unsupervised learning; if the environment is misrepresented, the “alertness” will be automatically increased to quickly identify the object.

5. The capacity is not limited by the number of input channels, and the stored objects do not need to be orthogonal.

ART Disadvantages

It is worth noting that the results of Fuzzy ART and ART 1 are heavily dependent on the order in which the training data are processed. This effect can be reduced to some extent by using a lower learning rate, but it remains regardless of the size of the dataset.

ART versions and differences

ART 1: A master-slave algorithm with a parallel architecture that uses set operations in the activation and matching functions of the algorithm. It mainly deals with the recognition problem of images containing only 0s and 1s (i.e. black and white).

ART 2: Can process grayscale (i.e. analog value) input.

ART 3: It has a multi-level search architecture, which combines the functions of the first two structures and expands the two-layer neural network to an arbitrary multi-layer neural network, and incorporates the bioelectrochemical reaction mechanism into the operation model of the neuron, so its functions are further expanded.