On dimensionality and sample size in statistical pattern classification

Kanal, Laveen N.; Chandrasekaran, B.

doi:10.1016/0031-3203(71)90013-6

Cited by 166 publications

(58 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It has been stated [4] , [7] that for the multivariant Gaussian case (with U k (i) and k σ denoting, respectively, the mean and variance of k in class i) with variances assumed equal for both classes, then if the variables are independent and…”

Section: Consequently Equation (18) Is Reduced Tomentioning

confidence: 99%

Optimal Projection From N-Dimensional Pattern Space Into a Plane

Hassan¹

2009

JNUS

View full text Add to dashboard Cite

This paper presents a discriminant algorithm that seeks to separate different classes as much as possible for discriminant analysis or dimension reduction. The optimization is achieved through the maximization of the Fisher ratio (which is defined as the ratio of the between-class scatter to the sum of within-class scatters).This algorithm for feature extraction shows improvement over the conventional feature selection algorithms used in remote sensing as well with other applications. The conducted experiments are accomplished using both simulated Gaussian and real airborne MSS/TM satellite data for both large and small sample size. Although the conducted experiments are performed over the case of two classes, extension to n-dimensions can be easily obtained using the binary decision tree.

show abstract

Section: Consequently Equation (18) Is Reduced Tomentioning

confidence: 99%

Optimal Projection From N-Dimensional Pattern Space Into a Plane

Hassan¹

2009

JNUS

View full text Add to dashboard Cite

show abstract

“…The value t (see section 2.2.2) is determined according to bibliographic research [16,12,13,11], in which the optimal dimension size depends on the experimental part in combination with the LVQ algorithms. These algorithms typically operate to preserve neighbourhoods on a network of nodes which encode the feature vector.…”

Section: Feature Vector Extractionmentioning

confidence: 99%

“…In the scientific practice the ideal size of learning feature vector of an artificial neural network it has been determined after experimentation and concretely from the minimization of training error procedure. Thus a size of 24 elements has been showed as an optimal size [13].…”

Section: Feature Vector Extractionmentioning

confidence: 99%

Specific Selection of FFT Amplitudes from Audio Sports and News Broadcasting for Classification Purposes

Poulos¹,

Bokos²,

Kanellopoulos³

et al. 2007

JGAA

View full text Add to dashboard Cite

In this paper we investigate the problem of classification between sports and news broadcasting. We detect and classify files that consist of speech and music or background noise (news broadcasting), and speech and a noisy background (sports broadcasting). More specifically, this study investigates feature extraction and training and classification procedures. We compare the Average Magnitude Difference Function (AMDF) method, which we consider more robust to background noise, with a novel proposed method. This method uses several spectral audio features which may be considered as specific semantic information. We base the extraction of these features on the theory of computational geometry using an Onion Algorithm (OA). We tested the classification procedure as well as the learning ability of the two methods using a Learning Vector Quantizer One (LVQ1) neural network. The results of the experiment showed that the OA method has a faster learning procedure, which we characterise as an accurate feature extraction method for several audio cases.

show abstract

“…This results in the so-called small sample size (SSS) problem, which is known to have significant influences on the performance of a statistical pattern recognition system (see e.g. [3,5,9,12,13,16,21,33,34]). …”

Section: Introductionmentioning

confidence: 99%

Kernel Discriminant Learning with Application to Face Recognition

Venetsanopoulos

Rogers

2005

Support Vector Machines: Theory and Applications

View full text Add to dashboard Cite

Abstract. When applied to high-dimensional pattern classification tasks such as face recognition, traditional kernel discriminant analysis methods often suffer from two problems: 1) small training sample size compared to the dimensionality of the sample (or mapped kernel feature) space, and 2) high computational complexity. In this chapter, we introduce a new kernel discriminant learning method, which attempts to deal with the two problems by using regularization and subspace decomposition techniques. The proposed method is tested by extensive experiments performed on real face databases. The obtained results indicate that the method outperforms, in terms of classification accuracy, existing kernel methods, such as kernel Principal Component Analysis and kernel Linear Discriminant Analysis, at a significantly reduced computational cost.

show abstract

On dimensionality and sample size in statistical pattern classification

Cited by 166 publications

References 8 publications

Optimal Projection From N-Dimensional Pattern Space Into a Plane

Optimal Projection From N-Dimensional Pattern Space Into a Plane

Specific Selection of FFT Amplitudes from Audio Sports and News Broadcasting for Classification Purposes

Kernel Discriminant Learning with Application to Face Recognition

Contact Info

Product

Resources

About