Information theory and signal analysis offer the possibility of identifying content-carrying structures within communications streams. However, the architecture of such structures (their inter-relation, potential context, and content), its semantic information, remains difficult to unpack. Using deep learning algorithms and newly-developed mathematical tools, analytical and theoretical physicist James Crutchfield from the University of California, Davis proposes to extract semantic information from acoustical databases of whale and dolphin vocalizations. This approach does not, a priori, promise to extract the meaning of such communication. Rather, it seeks to build a robust picture of the units of analysis (meaning-carrying information blocks and semantically-informational linkages) upon which such an ultimate exploration of meaning would have to be based. The project will draw from a variety of established databases and involve collaboration between Crutchfield and a designated computational team. This project falls under the TWCF Diverse Intelligences initiative.
Recurrent neural networks are used to forecast time series in finance, climate, language, and from many other domains. Reservoir computers are a particularly easily trainable form of recurrent neural network. Recently, a “next-generation” reservoir computer was introduced in which the memory trace involves only a finite number of previous symbols. The inherent limitations of finite-past memory traces are explored in this proposal.
