Tuba Ayhan

AN APPLICATION OF SMALL-WORLD CELLULAR NEURAL NETWORKS ON ODOR CLASSIFICATION

Many biological networks are constructed with both regular and random connections between neurons. Bio-inspired systems should prevent this mixed topology of biological networks while the artificial system is still realizable. In this work, a bio-inspired network which has many analog realizations, Cellular Neural Network (CNN) is investigated under existing random connections in addition to its regular connections: Small-World Cellular Neural Network (SWCNN). Antennal Lobe, an organ in the olfaction system of insects, is modeled with SWCNN by extending the network with the use of two types of processors on the same network. The model combined with a classifier, SVM and overall system is tested with a five-class odor classification problem. While all neurons are connected to each other with direct or indirect connections in CNNs, the idea of short-cuts does not provide an improvement in classification performance but the results show that the fault tolerance ability of SWCNN is better than the classical CNN.

Randomly reconfigurable Cellular Neural Network

Biological networks involve both regular and random connections. Moreover they employ more than one type of cells. Being widely used in bio-inspired systems, Cellular Neural Networks are practical to implement large networks due to their regularly defined connections between unit processors. However, this perfect regularity of the structure does not always match with applications. Although it is widely selected for retina like demonstrations itself, there is an absence of CNN for using it in other bio-inspired systems: an ordinary CNN has only one type of unit processor in one layer. However, sensory data processing in nature mainly depend on the collaboration of distinct dynamics. Neural mass models are suggested to mimic the joint effort of distinct types of neurons and they are widely used to simulate and understand brain activity. The regularity of an ordinary CNN benefits in implementation of the network. While protecting this simplicity, in this work, we propose a method to build a single layer cellular neural network that can perform a Wilson-Cowan like neural population model.

A Power Efficient System Design Methodology Employing Approximate Arithmetic Units

In this work a power efficient approximate system design methodology is introduced and its performance is demonstrated by a 2D-DCT implementation on Spartan 3 FPGA. The method is applicable to any system with arithmetic computation regardless of their architecture, because it utilizes the existing approximate arithmetic units. The novelty of the proposed method is its system analysis approach starting from the highest level and exploring through the sub-blocks down to the basic arithmetic units. It first evaluates a given system block diagram and sets the desired performance limits of each processing block to achieve the desired ultimate quality metric. Then, the arithmetic power consumption is minimized by employing the appropriate arithmetic units which are chosen by linear/non-linear programming with linear constraint solver. The tests on 2D-DCT implementation show a power reduction of 8% for a 0.01 dB PSNR loss for 128x128 images, on the average.

Supression of noise from speech signal using spectral subtraction with FPGA

Noise suppression from speech signal is one of the fundamental problems in audio processing applications. In this paper, a frequency domain based noise suppression algorithm is realized on FPGA. The algorithm is suitable for small, single microphone systems such as the hearing aids where the microphone and the speaker are very close to each other. Estimation of noise is performed adaptively so that the system becomes robust to changes in the environment.

A feature filtering method for eeg data classification

In this paper, a feature filtering algorithm for brain-computer interface which includes classification of EEG data is proposed. By this method, the features are evaluated according to a criterion based on the Mahalanobis distance between the classes. For some EEG data classification problems, the problem may be determining the features to be extracted, however for the problem of distinguishing between right, left and forward movement imagination, the features that most benefits in classification cannot be determined beforehand. Therefore, features are selected method from a set of all possible features by the proposed filtering to increase the performance and speed of the classifier.

Temperature‐Centric Evaluation of Sensor Transients

Controllable sensing conditions provide the means for diversifying sensor response and achieving better selectivity. Modulating the sensing layer temperature of metal‐oxide sensors is a popular method for multiplexing the limited number of sensing elements that can be employed in a practical array. Time limitations in many applications, however, cannot tolerate an ad‐hoc, one‐size‐fits‐all modulation pattern. When the response pattern is itself non‐stationary, as in the transient phase, a temperature program also becomes infeasible. We consider the problem of determining and tuning into a fixed optimum temperature in a sensor array. For this purpose, we present an empirical analysis of the temperature’s role on the performance of a metal‐oxide gas sensor array in the identification of odorants along the response transient. We show that the optimal temperature in this sense depends heavily on the selection of (i) the set of candidate analytes, (ii) the time‐window of the analysis, (iii) the feature extract...

Using a cellular neural network based olfactory bulb model for choosing the best sensor temperature for an odor classification problem

In this paper, a part of mamal olfaction system, olfactory bulb, is modelled by a Cellular Ceural Network and the performance of the model in an odor classification problem is evaluated for different sensör temperatures in order to figure out in which sensör temperature the most distinguishable data is recorded. The relevant probem in odor classification task is the slowy changing time response of the odor sensors and the model presented in this work is a structure that can be used to speed up odor processing.