Learning Equivalent Input Channel Mappings And Generalized Features For Pattern Transfer

Abstract

In many real-world modeling applications, it is needed to detect the origin of the patterns of the input data in addition to find the patterns themselves. Having the input data generated by a systematically organized set of input channels is very common in these applications. These input channels might also be of the same type. Therefore, the same pattern might be observed on different sets of input channels of the data, while it is caused by the one source in different localities. Sparse coding is a very powerful method for learning high-level patterns (i.e. high-level features) from raw data input. It is capable of learning an overcomplete basis which has the capacity to capture robust and discriminative patterns within the data. However, like many other feature learning algorithms, it is unable to detect identical features or stimuli on different sets of input channels. In this work, we propose a novel method to build general features that can be applicable to different sets of channels. This succinct representational model will express the stimuli independent of the locality in which they appeared. Simultaneously, different equivalent localities (equivalent channel sets) will be detected. As a result, when a feature is recognized on a channel set it can be transferred to the other equivalent channel sets. This enables the method to model and represent a pattern on localities where it has never been observed before.