Abstract
Recently, one of the standard discriminative training methods for pattern classifier design, i.e., Minimum Classification Error (MCE) training, has been revised, and its new version is called Large Geometric Margin Minimum Classification Error (LGM-MCE) training. It is formulated by replacing a conventional misclassification measure, which is equivalent to the so-called functional margin, with a geometric margin that represents the geometric distance between an estimated class boundary and its closest training pattern sample. It seeks the status of the trainable classifier parameters that simultaneously correspond to the minimum of the empirical average classification error count loss and the maximum of the geometric margin. Experimental evaluations showed the fundamental utility of LGM-MCE training. However, to increase its effectiveness, this new training required careful setting for hyperparameters, especially the smoothness degree of the smooth classification error count loss. Exploring the smoothness degree usually requires many trial-and-error repetitions of training and testing, and such burdensome repetition does not necessarily lead to an optimal smoothness setting. To alleviate this problem and further increase the effect of geometric margin employment, we apply in this paper a new idea that automatically determines the loss smoothness of LGM-MCE training. We first introduce a new formalization of it using the Parzen estimation of error count risk and formalize LGM-MCE training that incorporates a mechanism of automatic loss smoothness determination. Importantly, the geometric-margin-based misclassification measure adopted in LGM-MCE training is directly linked with the geometric margin in a pattern sample space. Based on this relation, we also prove that loss smoothness affects the production of virtual samples along the estimated class boundaries in pattern sample space. Finally, through experimental evaluations and in comparisons with other training methods, we elaborate the characteristics of LGM-MCE training and its new function that automatically determines an appropriate loss smoothness degree.
Similar content being viewed by others
Notes
While α y ’s are set common to all classes in most MCE embodiments, we also treat the case of different α y ’s for each class in consideration of the introduction of the automatic loss smoothness determination described in Section 3.
Even in the linear discriminant function case, D y (x; Λ) is not just a constant multiple of d y (x; Λ) because denominator ∥∇ x d y (x; Λ)∥ depends on trainable parameter set Λ.
LGM-MCE training is also applicable to other spaces than original input space 𝒳. Indeed, some of the authors of this paper recently formulated a Kernel MCE method that applied LGM-MCE training to kernel-based high-dimensional space and experimentally demonstrated its utility [23].
This relation is described in Chapter 4 of [7], too.
For large-scale applications, the CVML method may not be practical since it repeats \(O(N_{y}^{2})\) computation. In such cases, applying only the IQR-based method, which is just O(N y ) and requires no iterations, may be recommended.
As discussed previously, the term “approximately” can be omitted for linear discriminant functions or Euclidean-distance-based prototype classifiers even if \(\mathbf {\boldsymbol{x}}_{k}^{y}\) is not close to the boundary.
Bishop revealed that the minimization of the sum-of-square (or sum-of-cross-entropy) error with random noise added to the input data (not the transformed data points on misclassification measure space) is almost equivalent to the minimization of regularized error without noise [29]. Similarly, our approach, which produces virtual samples in the input pattern space (Fig. 5), may also minimize the regularized classification error without virtual samples. Theoretical analysis of this interesting issue remains future work.
The computation of best prototype p j assumes that a smooth “soft-max” function is approximated by the standard “max” operator for computational simplicity.
References
Jebara, T. (2004). Machine learning: Discriminative and generative. Kluwer.
Schlüter, R., Macherey, W., Müller, B., Ney, H. (2001). Comparison of discriminative training criteria and optimization methods for speech recognition. Speech Communication, 34, 287–310.
He, X., Deng, L., Chou, W. (2008). Discriminative learning in sequential pattern recognition. IEEE Signal Processing Magazine, 14–36.
Jiang, H. (2010). Discriminative training of HMMs for automatic speech recognition: a survey. Computer Speech & Language, 24(4), 589–608.
Juang, B.-H., & Katagiri, S. (1992). Discriminative learning for minimum error classification. IEEE Transactions Signal Processing, 40(12), 3043–3054.
Katagiri, S., Juang, B.-H., Lee, C.-H. (1998). Pattern recognition using a family of design algorithms based upon the generalized probabilistic descent method. Proceedings of the IEEE, 86(11), 2345–2373.
Bishop, C.M. (2006). Pattern recognition and machine learning. New York: Springer-Verlag.
Duda, O.H., & Hart, P. (1973). Pattern classification and scene analysis. New York: Wiley.
Fukunaga, K. (1990). Introduction to statistical pattern recognition, 2nd edn. San Diego: Academic Press.
Vapnik, V.N. (1995). The nature of statistical learning theory. New York: Springer-Verlag.
Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines. Cambridge University Press.
Abe, S. (2005). Support vector machines for pattern classification. London: Springer-Verlag.
He, T., Hu, Y., Huo, Q. (2007). An approach to large margin design of prototype-based pattern classifiers. Proceedings of the ICASSP, 2, II-625–628.
He, T., & Huo, Q. (2008). A study of a new misclassification measure for minimum classification error training of prototype-based pattern classifiers. Proceedings of the ICPR, 1–4.
Wang, Y., & Huo, Q. (2010). Sample-separation-margin based minimum classification error training of pattern classifiers with quadratic discriminant functions. Proceedings of the ICASSP, 1866–1869.
Watanabe, H., Katagiri, S., Yamada, K., McDermott, E., Nakamura, A., Watanabe, S., Ohsaki, M. (2010). Minimum error classification with geometric margin control. Proceedings of the IEEE, 2170–2173.
Watanabe, H., & Katagiri, S. (2011). Minimum classification error training with geometric margin enhancement for robust pattern recognition. Proceedings IEEE MLSP (CD version), 1–6.
Freund, Y., & Schapire, R.E. (1997). A decision-theoretic generalization of on-line learning and an application to boosting. Soviet Journal of Computer and Systems Sciences, 55(1), 119–139.
Liu, C., Jiang, H., Rigazio, L. (2006). Recent improvement on maximum relative margin estimation of HMMs for speech recognition. Proceedings of the ICASSP, 1, I-269–272.
Jiang, H., Li, X., Liu, C. (2006). Large margin hidden Markov models for speech recognition. IEEE Transactions on Audio, Speech, Language Processing, 14(5), 1584–1595.
Li, J., Yuan, M., Lee, C.-H. (2007). Approximate test risk bound minimization through soft margin estimation. IEEE Transactions on Audio, Speech, Language Processing, 15(8), 2393–2404.
Yu, D., Deng, L., He, X., Acero, A. (2008). Large-margin minimum classification error training: a theoretical risk minimization perspective. Computer Speech & Language, 22, 415–429.
Tanaka, H., Watanabe, H., Katagiri, S., Ohsaki, M. (2012). Experimental evaluation of kernel minimum classification error training. Proceedings of the IEEE TENCON, 1–6.
Watanabe, H., Tokuno, J., Ohashi, T., Katagiri, S., Ohsaki, M. (2011). Minimum classification error training with automatic setting of loss smoothness. Proceedings of the IEEE MLSP, 1–6.
McDermott, E., & Katagiri, S. (2004). A derivation of minimum classification error from the theoretical classification risk using Parzen estimation. Computer Speech & Language, 18, 107–122.
Ohashi, T., Tokuno, J., Watanabe, H., Katagiri, S., Ohsaki, M. (2011). Automatic loss smoothness determination for large geometric margin minimum classification error training. Proceedings IEEE TENCON, 260–264.
Ohashi, T., Watanabe, H., Tokuno, J., Katagiri, S., Ohsaki, M., Matsuda, S., Kashioka, H. (2011). Increasing virtual samples through loss smoothness determination in large geometric margin minimum classification error training. Proceedings IEEE ICASSP, 2081–2084.
Silverman, B.W. (1986). Density estimation for statistics and data analysis. Chapman & Hall/CRC.
Bishop, C.M. (1995). Training with noise is equivalent to Tikhonov regularization. Nural Computation, 7, 108–116.
Sato, A., & Yamada, K. (1996). Advances in neural information processing systems (Vol. 8, pp. 23–429). MIT Press.
Sato, A., & Yamada, K. (1998). A formulation of learning vector quantization using a new misclassification measure. Proceedings ICPR1998, 322–325.
Crammer, K., & Singer, Y. (2001). On the algorithmic implementation of multiclass kernel-based vector machines. Journal Machine Learning Research, 2, 265–292.
Author information
Authors and Affiliations
Corresponding author
Additional information
This work was supported in part by Grant-in-Aid for Scientific Research (B), No. 22300064.
Rights and permissions
About this article
Cite this article
Watanabe, H., Ohashi, T., Katagiri, S. et al. Robust and Efficient Pattern Classification using Large Geometric Margin Minimum Classification Error Training. J Sign Process Syst 74, 297–310 (2014). https://doi.org/10.1007/s11265-013-0760-4
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11265-013-0760-4