Home

People

Projects

Research topics

Publications

Contact		 Hybrid Model Learner (HML)

  • A special architecture for model learning.
  • This new discovery* combines linear and nonlinear operations and utilizes efficient statistical methods to realize noise tolerant, unsupervised pattern discovery and learning by modeling regularities in sequences.
  • The learned pattern models are related to context-free grammars.
      * The discovery is based on the work performed during 2002-2009 by professor Unto K. Laine. Patent is pending (FI-20095708)
    • HML is hybrid in many dimensions.

    HML has many different functionalities integrated

    HML general architecture


    Noise or Pattern?

  • Based on its internal statistical models and methods HML can separate noise from perceptually relevant, structural information based solely on the information provided by the input sequence under study.
  • The only prior, very general knowledge how to separate patterns from noise, is given by the internal statistical models and methods of the HML.


    • D = input sequence
    M = model for D
    G = generator/extractor
    D' = residual sequence

    1° The statistical properties of the input sequence D are analyzed in order to create models M(D) for it.
    2° The models are used to remove the structures found from the input sequence (extractor G).
    3° A compressed residual sequence D' is created.

    HML experiments and demonstrations

    Example 1: Compression of binary sequences
    • HML performs about 30% better than GZIP
    Example 2: Classification of pseudorandom sequences
    • Sequences generated by standard pseudorandom generators can be recognized and separated from the natural random sequences.
    • Weak, hidden patterns in the pseudorandom sequences can be detected and learned by the HML.
    Example 3: A nonlinear, high-resolution HML filter
    • dfdt ~ = 0.1 (classical limit in linear case = 0.5).
    • The obtained result is close to that measured from human auditory system.
    Example 4: Compression of 10 000 first digits of pi
    • HML expands the sequence less than any other known lossless compression method.
    Example 5: Patterns in genomic sequences
    • HML is able to find important subsequences without any external help entirely based on its internal mechanisms.
    Example 6: Word recognition in continuous speech
    • HML is able to learn word models just by exposing it to sentences having the words (supervised learning).
    • After the learning phase, HML is ready to recognize similar words in a continuous speech.
    Publications on the theory and simulations of HML are under preparation (6/09)

    Compression of binary sequences where the probability of 1 is q and that of 0 is (1-q). HML tested with two different preprocessing.




      A Chinese proverb:

        Men who say it cannot be done, should not interrupt those doing it.

    Speech Technology Team, 2009