Bio-Inspired Techniques in Human-Computer Interface for Control of Assistive Devices: Bio-Inspired Techniques in Assistive Devices

Introduction

Pattern recognition is a method of identifying the input information into particular category/class from various classes. Various researches have been carried out in improving control of intelligent assistive devices in the various stages of pattern recognition techniques, namely data preprocessing, feature extraction, feature selection/reduction, classification along with the development of control strategy of electric motor. In pattern recognition, the data usually considered as the raw measurements or raw values taken from the subjects to be classified. A simple block diagram of pattern recognition based control of assistive devices is shown in Figure 1.

Figure 1.

Block diagram of pattern recognition based control of assistive device

The term feature in pattern recognition, refer to the result of the transformations applied to the raw data in order to transform them into another domain or space using time domain/ frequency domain/ time-frequency domain technique. Although, many features can be extracted from raw data for decoding intention and not all of them possesses discriminant capabilities. Some of the extracted features could cause confusion and degrade the classifier. Further, smaller the dimension of the feature vector, lesser the computation time and memory requirements. Therefore, choice of features or reduction of features is essential. Feature dimension reduction provides a method to decide whether it is necessary to include more features that would significantly contribute to the performance of the classifier. It is not a trivial to select the best set of features or the best transformation. The features must be selected or transformed based on the given problem. For the feature selection, some neural networks, population based bio-inspired techniques can be used. The features selection process involves choice of subset of extracted features in feature space by starting with all/without features or subset consisting of random features from the feature space. Feature selection process in the context of assistive devices will be reviewed later in this chapter. The Fourier transform and time-frequency transom yield coefficients of larger dimension and few of its coefficients carry the useful information to obtain the good classification performance. In literature, researchers applied feature reduction using linear or nonlinear projection of features to transform high dimensional feature space to lower dimensional feature space. A very popular method of feature reduction is principal component analysis (PCA) in which the features are projected to lower dimensional space to visualize the underlying class by linear projection. There are other many feature reduction approaches like linear discriminant analysis (LDA), fuzzy discriminant analysis (FDA), self-organizing feature map (SOFM), kernel-based FDA, etc. are discussed in literatures.

Background

The feature selection aims to cut the dimensionality by eliminating irrelevant and redundant features, thus considering a subset of features that characterize the best discrimination of patterns. Since, it is difficult to find the discriminant features and selection algorithm using population based technique, use evaluation function to find the features of good discriminating capability. Feature selection would ease the problem of over fitting and reduces the classifier computation time. The existing feature selection techniques in the literature are divided as filters and wrappers according to their dependency on the classification algorithms. Typically, feature selection techniques could be useful in supervised or unsupervised learning algorithms. The filter approach is independent of classification algorithm and uses statistical properties to identify the relevant features. Due to this capability, filter approach is computationally preferred than the wrapper approach. However, it has disadvantage of local optimal solution due to its single iteration. The wrapper approach is based on a classification algorithm need more computation time, but more accurate than the filter approach. The hybrid approach combines the advantages of the filter and wrapper.