This paper presents a novel incremental feature selection method based on minimum uncertainty and feature sample elimination (referred as MUSE). Feature selection is an important step in machine learning. In an incremental feature selection approach, past approaches have attempted to increase class relevance while simultaneously minimizing redundancy with previously selected features. One example of such an approach is the feature selection method of minimum Redundancy Maximum Relevance (mRMR). The proposed approach differs from prior mRMR approach in how the redundancy of the current feature with previously selected features is reduced. In the proposed approach, the feature samples are divided into a pre-specified number of bins; this step is referred to as feature quantization. A novel uncertainty score for each feature is computed by summing the conditional entropies of the bins, and the feature with the lowest uncertainty score is selected. For each bin, its impurity is computed by taking the minimum of the probability of Class 1 and of Class 2. The feature samples corresponding to the bins with impurities below a threshold are discarded and are not used for selection of the subsequent features. The significance of the MUSE feature selection method is demonstrated using the two datasets: arrhythmia and hand digit recognition (Gisette), and datasets for seizure prediction from five dogs and two humans. It is shown that the proposed method outperforms the prior mRMR feature selection method for most cases. For the arrhythmia dataset, the proposed method achieves 30 percent higher sensitivity at the expense of 7 percent loss of specificity. For the Gisette dataset, the proposed method achieves 15 percent higher accuracy for Class 2, at the expense of 3 percent lower accuracy for Class 1. With respect to seizure prediction among 5 dogs and 2 humans, the proposed method achieves higher area-under-curve (AUC) for all subjects.

This paper presents novel architectures for machine learning based classifiers using stochastic logic. Two types of classifier architectures are presented. These include: linear support vector machine (SVM) and artificial neural network (ANN). Stochastic computing systems require fewer logic gates and are inherently fault-tolerant. Thus, these structures are well suited for nanoscale CMOS technologies. These architectures are validated using seizure prediction from electroencephalogram (EEG) as an application example. To improve the accuracy of proposed stochastic classifiers, a novel approach based on linear transformation of input data is proposed for EEG signal classification using linear SVM classifiers. Simulation results in terms of the classification accuracy are presented for the proposed stochastic computing and the traditional binary implementations based datasets from one patient. Compared to conventional binary implementation, the accuracy of the proposed stochastic ANN is improved by 5.89%. Synthesis results are also presented for EEG signal classification. Compared to the traditional binary linear SVM, the hardware complexity, power consumption and critical path of the stochastic implementation are reduced by 78%, 74% and 53%, respectively. The hardware complexity, power consumption and critical path of the stochastic ANN classifier are reduced by 92%, 88% and 47%, respectively, compared to the conventional binary implementation.