We propose to apply the Tree Structured Vector Quantizer (TSVQ) [.Baras Wolk 1993.] to the multiscale representation of spectra of selected databases. Our goal is to isolate and organize characteristic spectral features of the database. This process would proceed by first computing the multiscale representation of a large number of spectral profiles from a selected database, such as signals from acoustic sensor arrays, or underwater transients. Then, at each resolution, this signal space is partitioned into different clusters or cells which are determined by repeated application of the Linde-Buzo-Gray (LBG) algorithm [.baras Wolk 1993.]. The algorithm is first applied to the coarsest resolution of the data vectors, which results in relatively few clusters or classes. The interpretation of these clusters likely depends on the nature of the database. For instance, for underwater transients and battlefield acoustic sensor signals the clusters may distinguish broad classes of vessels or vehicles ( Thrust area V(B)). For speech vowels, they may signify fundamental vocal tract configurations such as the open-closed, or front-back distinctions ( Thrust area III(D)).
However, it is the clustering performed next at the finer resolution which yields the newest insights. Specifically, each cluster (or equivalence class of coarse representations) is split next with another round of repeated applications of the LBG algorithm. Thus, a tree-structure emerges at which each new layer corresponds exactly to partitions based on the next finer resolution data. Rules and criteria for performing this partitioning have already been developed and applied in engineering tasks such as the analysis of ship radar-returns [.baras wolk 1993.].
Finally, we plan to exploit another innovation in this and subsequent algorithms. It concerns the criteria by which the clustering at each scale occurs. Specifically, many of these algorithms can be adaptive in nature to reflect changes in the data, or to balance competing demands of the application. For instance, vector quantization can be viewed both as a compression and as a classification algorithm (as we have been emphasizing), and hence one can measure its performance by satisfying multi-objective functions, e.g., incorporating the Bayes risk (measuring classification performance) in the distortion measure (measuring compression performance). An efficient way of achieving this goal is represented by the Learning Vector Quantization (LVQ) algorithm of Kohonen [. Kohonen 1990.], an unsupervised neural network classifier which we have used and analyzed extensively [.baras 1990.].