A modified RBF neural network for efficient current-mode VLSI implementation

2013 4th International Symposium on Electrical and Electronics Engineering (ISEEE)

Florea

2013

This paper presents a study of three classification methods applied on natural images with the goal of detecting the eye regions. The challenge that we aim to solve is finding the eyes on faces expressing emotions -this is extremely difficult because the shape of the facial features changes drastically when the subject goes through an emotional state. We attempt to solve this challenge by looking for eye patterns in an image and categorizing them using neural network classification approaches. The image features used are integral projections encoded by TESPAR method, thus reducing the dimensionality of the problem to vectors of 60 elements -we observed that the network classification ability becomes more robust on input samples computed in this manner and the computational complexity is lowered, such that the system gets closer to making real-time decisions. The paper presents experimental results of three classifiers (SVM, Naïve Bayes and MLP) obtained on the Cohn-Kanade reference database. The measurements show that by tuning each classifier properly and by performing a postprocessing step, we can obtain accuracies of over 90%.

Section: B Modified Naïve Bayes Classifiermentioning

confidence: 99%

A comparison of several classifiers for eye detection on emotion expressing faces

Boia¹,

2013 4th International Symposium on Electrical and Electronics Engineering (ISEEE)

Florea

2013

“…The second modification comes as a consequence of a result in [4] where it was demonstrated that the exponential basis function…”

Section: B Simplified Variants Of the Nb Classifiermentioning

confidence: 99%

“…In this work, we connect Naïve Bayes theory to results in our previous works where a simplified neural model called a RBF-M network [4] was defined as an efficient and effective neural model that is suitable for embedded systems (i.e. with a relatively simple training algorithm and Gaussian basis function being replaced to a piecewise-linear formula in order to reduce the implementation complexity).…”

Section: Introductionmentioning

confidence: 99%

A modified Naïve Bayes classifier for efficient implementations in embedded systems

ISSCS 2011 - International Symposium on Signals, Circuits and Systems

2011

In this paper we propose two modifications of the Naïve Bayes (NB) algorithm, in order to reduce its complexity such that it may be effectively implemented with simple operators in embedded computing systems. A first modification is the introduction of a tuning parameter similar to the radius in radial basis function neural networks, it allows improving classification performance. The second modification is the approximation of exponential function with a piecewise-linear function that allows efficient implementation in embedded systems. Using a large set of benchmark problems, comparisons with "standard" NB and with other classifiers (such as SVM and a modified RBF) provided that modified NB learns very fast and may have a very efficient implementation providing a good accuracy.

“…In [6] a first Java desktop application for modeling and simulation of several novel neural network architectures with low complexity implementation [5][13] [14] was developed. These neural architectures are all kernel networks optimized for low complexity while maintaining performance comparable to the traditional neural classifiers.…”

Section: Introductionmentioning

confidence: 99%

“…The JLCNN library described in this paper implements five low complexity neural network architectures, namely RBF-M [5], Simplicial [13], SORT [14], neo-fuzzy neural networks [17], and the modified Naïve Bayes [9]. It can be easily extended to include new neural models while it includes unique functions to access the data files (e.g.…”

Section: Introductionmentioning

confidence: 99%

JLCNN: An object-oriented Java package for low complexity neural networks

2013 4th International Symposium on Electrical and Electronics Engineering (ISEEE)

2013

In this paper, a Java object-oriented package (abbreviated JLCNN) is presented. Its objective is to implement several low complexity neural networks architectures, for further use in various embedded systems applications. The main advantages of the JLCNN package are the rapid development of various applications based on low complexity neural architectures and the easiness in porting such neural architectures on various computing platforms. As any other object-oriented library it allows reducing the development time for a new software product and provides convenient code maintenance, reusability and portability. The library is also useful for research and educational purposes, as a basic kernel on which new modules can be easily added. For instance it may be expanded to image processing applications.