Cognitive Component Analysis

This dissertation concerns the investigation of the consistency of statistical regularities in a signaling ecology and human cognition, while inferring appropriate actions for a speech-based perceptual task. It is based on unsupervised Independent Component Analysis providing a rich spectrum of audio contexts along with pattern recognition methods to map components to known contexts. It also involves looking for the right representations for auditory inputs, i.e. the data analytic processing pipelines invoked by human brains. The main ideas refer to Cognitive Component Analysis, defined as the process of unsupervised grouping of generic data such that the ensuing group structure is well-aligned with that resulting from human cognitive activity. Its hypothesis runs ecologically: features which are essentially independent in a context defined ensemble, can be efficiently coded as sparse independent component representations. The focus has been to construct a preprocessing pipeline for COCA to search for the ‘cognitive structure’, and to measure the alignment of the resulting from unsupervised learning and human cognition. Based on the nature of human auditory system and psychoacoustics, we have constructed the pipeline: feature extraction; feature integration; energy based sparsification; and principal component analysis. To test whether human uses information theoretically optimal ICA methods in higher cognitive functions, is the main concern in this thesis. It is well-documented that unsupervised learning discovers statistical regularities. However human cognition is too complicated and not yet fully understood. Nevertheless, in our approach we represent human cognitive processes as a classification rule in supervised learning. Thus we have devised a testable protocol to test the consistency of statistical properties and human cognitive activity, i.e. unsupervised learning of perceptual inputs and supervised learning of inputs together with manually obtained labels. The comparison has been carried out at different levels. This protocol has successfully revealed the consistency of two classifications via several speech-based cognitive tasks.