動的類似度を用いたワンショット学習則

抄録

We have proposed the one-shot learning algorithm called SOLAR (Supervised One-shot Learning Algorithm for Real number inputs) which needs only a single presentation of training examples, and the learning speed is extremely fast. However, the generalization ability is not satisfactory. In this paper, we propose SOLAR 2 which has good generalization ability. To improve the generalization ability, we introduce the concept of Adaptive Similarity for grouping the training examples. Adaptive Similarity adapts the similarity measure on the course of training. The characteristics of SOLAR 2 are as follows: it learns extremely fast, constructs the network automatically, and it has no local minima problem.