14th International Congress of Phonetic Sciences (ICPhS-14) San Francisco, CA, USA August 1-7, 1999

On Error Criteria in Human Classification Modeling

Louis ten Bosch (1,2), Roel Smits (3)

(1) Lernout & Hauspie Speech Products N.V., Ieper, Belgium
(2) Institute of Phonetic Sciences/IFOTT, University of Amsterdam, NL
(3) Max Planck Institute, Nijmegen, The Netherlands

A number of error measures will be discussed that are used for training and testing classification algorithms designed to simulate human classification behaviour. These are (1) the error criterion based on multinomial decision strategy (average log likelihood ratio), (2) the mean squared error (MSE) based on the L2 (Euclidean metric), (3) an error criterion based on similarity, (4) a novel one, the average log likelihood ratio. We will not focus on particular minimalization methods that are inherent to specific numerical minimization schemes, such as the back propagation method, stochastic annealing, etc., but rather on the conceptual differences of the error measures mentioned.
   The classifiers are implemented by means of a feedforward network. The training will be considered to be supervised in all cases.

Full Paper

Bibliographic reference.  Bosch, Louis ten / Smits, Roel (1999): "On error criteria in human classification modeling", In ICPhS-14, 1949-1952.