画像意識用の項目出力

最近スタンフォードとミシガン大学の研究者が行った機会学習実験について学んでいます。The point is to extract a feature set (set of coordinates that are functions of the input data). What is interesting is that the feature extraction is carried out by a set of unsupervised learning algorithms, and the effectiveness of the feature extraction is evaluated using the support vector machine supervised learning algorithm. The unsupervised learning algorithms are k-means clustering, mixture of Gaussians, sparse auto-encoder, and Restricted Boltzmann Machine. Basically, these algorithms output a descriptive numerical vector for each image input which is of lower dimensional it than the original n-by-n pixel input data set. In the case of K-means, this vector indicates which of the K clusters' centroids the input image is nearest to. Note that the image feature vector is obtained by sampling overlapping patches from the image, computing a k-dimensional vector for each patch, then aggregating the patch vectors into an image feature vector.
An auto-encoder is a feed forward neural network, and a Restricted Boltzmann Machine is a two-layer recursive neural network.