An Algorithm for Two-Dimensional Pattern Detection by Combining Echo State Network-Based Weak Classifiers

We propose a two-dimensional pattern detection algorithm based on Echo State Network (ESN). In the field of pattern detection, the gradient descent method has often been used in training a multilayer perceptron or deep learning. The training rule of ESN is one-shot ridge regression, which is not compatible with gradient descent. To train an ESN network as a pattern detection classifier, we followed the basic training rule of ESN so that training input vectors are organized as a training input matrix, where each column vector corresponds to an original training image. Here we assumed that aligning a certain number of the same column vectors, which corresponds to the same training image, in the training input matrix will form a short-term memory or an attractor. We organized the input training matrix as such and trained the ESN network. We made sure that the output vector of the trained ESN network for each training input pattern is approaching to the ground-truth vector by repeatedly feeding the same training image to the ESN input layer. The resulting performance of a single ESN classifier is, however, relatively poor. Therefore, we used an ensemble classifier framework by combining multiple ESN weak classifiers. We show the basic performance of ensemble-based ESN pattern classification and discuss the remaining tasks to improve the performance of the ensemble-based ESN classifiers.