Finding a rich lower dimensional representation

Artificial neural networks are the key driver of progress in various semantic computer vision tasks such as age prediction and digit classification. For the successful application of neural network algorithms, the representation of the data is an important factor. A good representation can significantly simplify a regression or prediction task. For semantic computer vision tasks, such as digit classification, a neural network needs to be trained such that it can map images from the spatial representation or spatial domain to the class domain. Mapping between domains requires a complex neural network with multiple layers and a large amount of labeled training data spanning the the input and output domain. In order to reduce the amount of required labeled data, we propose to map the images to a rich lower dimensional representation which is correlated with the class domain