The SOLAR algorithm
Kosei Demura, M. Kajiura, Y. Anzai · 1994
We propose SOLAR (supervised one-shot learning algorithm for real number inputs). SOLAR requires only a single presentation of real number input data, so it can learn very quickly compared to the backpropagation algorithm (BP). We introduce a new similarity matrix which measures Euclidean distance of the training set. From the topology of the similarity matrix, the structure of network, learning parameters and linear threshold functions are determined. Since it uses the structural method, SOLAR is suited to the dynamic environment, e.g. addition or subtraction of training instances, input units and output units. The main contribution of this paper is that SOLAR can handle analog inputs, so it can be easily extended to learning temporal sequences and improve the ability of the generalization.>