Sparse Coding Approaches to Language Acquisition
One of the scientific challenges is the understanding of the efficiency, flexibility and intelligence of biological systems. Their ability to is their ability to learn, viz. aquire knowledge of the interaction with the environment, in order to use it later among the characteristics. In this context the priority programme pursues the problem of autonomy. Instead we pursue the aim of autonomous learning, viz. independent learning, independent collection of information about a complex environment and independent education about structured re–presentation and generalized models of what has been learned.
This project at the University of Paderborn aims at the development of one of the systems for learning of reference pattern for unsupervised learning of a language. The machine shall descover recurring patterns within the continous spoken entering speech signal and learn an inventory of units, namely on two different levels of abstraction: on the one hand on the level of sounds and on the other hand on the level of words. Procedures are supposed to be used from the area of sparse coding in order to find a representation of a speech signal which is fed by the presentation of the speech signal within the short–term spectral range by dint of a linear combination of basic vectors. While non–negative matrix factorization (NMF) has already been used upon language, there are other procedures which require the non–negativity of matrix elements, so they can be applied better for the common parameterization of speech signals, as the Mel-Frequency Cepstral Coefficients. A promising procedure is the k-singular value decomposition (k-SVD), which so far has been primarily applied in computer vision. All of these learning procedures have to be extended, so they can add to the learning of typical spectral patterns and gather temporal correlation of speech signals. In addition approaches out of the area of dynamic time adjustment and speech recognition based on the "hidden" Markov–model are applied. On the first, lower level of the decomposition of the entering speech signal recurring sound units shall be discovered. On the second, higher level of abstraction word or phrases units similar to the procedure on the first level are learned, based upon a description of the lower levels with the help of n-grams, viz. with the help of the frequency of sound units. The lower level shall hand down the posteriori probabilities to the higher level, so an premature definite decision concerning the sounds can be avoided.
This project is one of 15 projects in the SPP "Autonomous Learning" of the DFG.