Panu Somervuo

Self-Organizing Maps of Symbol Strings

Abstract Unsupervised self-organizing maps (SOMs), as well as supervised learning by Learning Vector Quantization (LVQ) can be defined for string variables, too. Their computing becomes possible when the SOM and the LVQ algorithms are expressed as batch versions, and when the average over a list of symbol strings is defined to be the string that has the smallest sum of generalized distance functions from all the other strings.

How to make large self-organizing maps for nonvectorial data

The self-organizing map (SOM) represents an open set of input samples by a topologically organized, finite set of models. In this paper, a new version of the SOM is used for the clustering, organization, and visualization of a large database of symbol sequences (viz. protein sequences). This method combines two principles: the batch computing version of the SOM, and computation of the generalized median of symbol strings.

Self-Organizing Maps and Learning Vector Quantization forFeature Sequences

The Self-Organizing Map (SOM) and Learning Vector Quantization (LVQ) algorithms are constructed in this work for variable-length and warped feature sequences. The novelty is to associate an entire feature vector sequence, instead of a single feature vector, as a model with each SOM node. Dynamic time warping is used to obtain time-normalized distances between sequences with different lengths. Starting with random initialization, ordered feature sequence maps then ensue, and Learning Vector Quantization can be used to fine tune the prototype sequences for optimal class separation. The resulting SOM models, the prototype sequences, can then be used for the recognition as well as synthesis of patterns. Good results have been obtained in speaker-independent speech recognition.

Bird song recognition based on syllable pair histograms

Bird song can be divided into a sequence of syllabic elements. We investigate the possibility of bird species recognition based on the syllable pair histogram of the song. This representation compresses the variable-length syllable sequence into a fixed-dimensional feature vector. The histogram is computed by means of Gaussian syllable prototypes which are automatically found given the song data and the dissimilarity measure of syllables. Our representation captures the use of the syllable alphabet and also some temporal structure of the song. We demonstrate the method in bird species recognition with song patterns obtained from fifty individuals belonging to four common passerine bird species.

Classification of the harmonic structure in bird vocalization

The article is related to the development of techniques for automatic recognition of bird species by their sounds. It has been demonstrated earlier that a simple model of one time-varying sinusoid is very useful in classification and recognition of typical bird sounds. However, a large class of bird sounds are not pure sinusoids but have a clear harmonic spectrum structure. We introduce a way to classify bird syllables into four classes by their harmonic structure.

Experiments with linear and nonlinear feature transformations in HMM based phone recognition

Feature extraction is the key element when aiming at robust speech recognition. Both linear and nonlinear data-driven feature transformations are applied to the logarithmic mel-spectral context feature vectors in the TIMIT phone recognition task. Transformations are based on principal component analysis (PCA), independent component analysis (ICA), linear discriminant analysis (LDA) and multilayer perceptron network based nonlinear discriminant analysis (NLDA). All four methods outperform the baseline system which consists of the standard feature representation based on MFCCs (mel-frequency cepstral coefficients) with the first-order deltas, using a mixture-of-Gaussians HMM recognizer. Further improvement is gained by forming the feature vector as a concatenation of the outputs of all four feature transformations.

Clustering and Visualization of Large Protein Sequence Databases by Means of an Extension on the Self-Organizing Map

New, more effective software tools are needed for the analysis and organization of the continually growing biological databases. An extension of the Self-Organizing Map (SOM) is used in this work for the clustering of all the 77,977 protein sequences of the SWISS-PROT database, release 37. In this method, unlike in some previous ones, the data sequences are not converted into histogram vectors in order to perform the clustering. Instead, a collection of true representative model sequences that approximate the contents of the database in a compact way is found automatically, based on the concept of the generalized median of symbol strings, after the user has defined any proper similarity measure for the sequences such as Smith-Waterman, BLAST, or FASTA. The FASTA method is used in this work. The benefits of the SOM and also those of its extension are fast computation, approximate representation of the large database by means of a much smaller, fixed number of model sequences, and an easy interpretation of the clustering by means of visualization. The complete sequence database is mapped onto a two-dimensional graphic SOM display, and clusters of similar sequences are then found and made visible by indicating the degree of similarity of the adjacent model sequences by shades of gray.

Online algorithm for the self-organizing map of symbol strings

In this work an online algorithm is presented for the construction of the self-organizing map (SOM) of symbol strings. Each node of the SOM grid is associated with a model string which is a variable-vector sequence. Smooth interpolation method is applied in the training which performs simultaneous adaptation of the symbol content and the length of the model string. The efficiency of the method is demonstrated by the clustering of a 100,000-word English dictionary.

Competing hidden Markov models on the self-organizing map

This paper presents an unsupervised segmentation method for feature sequences based on competitive-learning hidden Markov models. Models associated with the nodes of the self-organizing map learn to become selective to the segments of temporal input sequences. Input sequences may have arbitrary lengths. Segment models emerge then on the map through an unsupervised learning process. The method was tested in speech recognition, where the performance of the emergent segment models was as good as the performance of the traditionally used linguistic speech segment models. The benefits of the proposed method are the use of unsupervised learning for obtaining the state models for temporal data and the convenient visualization of the state space on the two-dimensional map.

Speech Dimensionality Analysis on Hypercubical Self-Organizing Maps

The problem of finding the intrinsic dimension of speech is addressed in this paper. Astructured vector quantization lattice, Self-Organizing Map (SOM), is used as a projection space for the data. The goal is to find a hypercubical SOM lattice where the sequences of projected speech feature vectors form continuous trajectories. The effect of varying the dimension of the lattice is investigated using feature vector sequences computed from the TIMIT database.