Zbigniew Szymański

Deep learning classifier for fall detection based on IR distance sensor data

The goal of research is the fall detection in elderly residents based on infra red depth sensor measurements. Our attention is focused on statistical properties as generalization. The effectiveness of discriminative statistical classifiers (multilayer perceptron) is improved by addition of feature selection block by Gram-Schmidt orthogonalization, which determines the ranking of the features, and NPCA block, which transforms the raw data into a nonlinear manifold and reduces the dimensionality of the data. Performance of our system measured in terms of sensitivity is 92% and precision is 93%, which means it can be used for real life applications.

Neural network classifier for fall detection improved by Gram-Schmidt variable selection

The paper describes results of research on the fall detection in elderly residents based on infra red depth sensor measurements. We present the methodology of data acquisition, preprocessing and the feature extraction. Multilayer perceptron is used for classification. In order to improve the classifier generalization feature selection block by Gram-Schmidt orthogonalization is added. It determines the ranking of the features and enables to reduce the dimensionality of the data. Performance of our system measured in terms of sensitivity is 92% and precision is 93%, which means it can be used for real life applications.

Reconstruction of environment model by using radar vector field histograms

The paper presents a method of creating an environment model in collision avoidance system for unmanned aerial vehicles (UAV). The environment model is generated by the procedures processing the data from on-board equipment and digital maps. The main sensor that provides information about the current situation around the UAV is a radar obstacle detector. Each detected object is defined by such parameters as distance, speed and the number of radial zone. The method is based on the idea of the certainty grid introduced in vector field histogram method which is used as a probabilistic representation of the obstacles. The tests of developed algorithm were performed in simulated environment.

Improved nearest neighbor classifier based on local space inversion

We are offering a method to improve classification performance of the k nearest neighbors algorithm when the neighborhood is large. Our method limits the number of considered neighbors from the training set by selecting only those samples which would be neighbors in the computed Voronoi diagram of the training set plus classified sample.

Prediction of protein phosphorylation sites using classification trees and SVM classifier

The paper presents a method of solving the problem of protein phosphorylation sites recognition. Six classifiers were created for prediction whether specified amino acid sequences represented as a 9-character strings react with given types of the kinase-enzymes. The method consists of three steps. Positions in the amino acid sequences significant for classification are found with the use of classification trees in the first step. Afterwards, the symbols composing the sequences are mapped to the real numbers domain using the Gini index method. The last step consists of creating the SVM classifiers as the final prediction models. The paper contains evaluation of the obtained results and the description of the methods applied to evaluate the quality of the classifiers.

Approximation of HRPITS results for SI GaAs by large scale support vector machine algorithms

For the first time large-scale support vector machine algorithms are used to extraction defect parameters in semi-insulating (SI) GaAs from high resolution photoinduced transient spectroscopy experiment. By smart decomposition of the data set the SVNTorch algorithm enabled to obtain good approximation of analyzed correlation surface by a parsimonious model (with small number of support vector). The extracted parameters of deep level defect centers from SVM approximation are of good quality as compared to the reference data.

Symmetric block encoder based on reversible circuits

The goal of the presented work is a project of a novel symmetric block encoder. The basic processing elements are cascades of reconfigurable reversible gates changing the type of gates depending on the encryption key. The presented solution proposes the use of sixteen 8-bit cascades, which configuration requires a 640-bit key. The input information is processed in five rounds. The encryption keys in the subsequent rounds differ. The design was modeled in VHDL language and placed in an FPGA chip. The project is scalable, i.e. depending on the needs, it can be modified by changing the number of gates in the cascade, the width of the information block being processed, which may result in an increase or decrease in the width of the encryption key. The number of rounds may also be modified. The large size of the encryption key should ensure the safety of the encrypted data.

Memristive CNN for trajectory recognition

The paper describes application of CNN with memristive units to simple movement recognition. Each cell of CNN comprised input and output template connected to voltage controlled memristor. The experimental results were obtained by using simulation software. The proposed model proved to be capable of recognizing simple pixel trajectories over a two-dimensional grid.

Implementation of reversible gates in FPGA structure

The paper presents reversible circuits modelling methods in the FPGA structures. Three different methods of reconfigurable reversible gate descriptions are proposed - a direct method of output calculation and two methods of composing the gate from Fredkin and Toffoli gates. The paper shows both 4-bit and 8-bit gates. Application of developed test bench verified the correctness of our designs. The presented descriptions can be used for modelling of complex elements used e.g. in cipher machines based on quantum algorithms.

Deep learning classifier based on NPCA and orthogonal feature selection

In this paper the idea of deep learning classifier is developed. The effectiveness of discriminative classifier, as e.g. multilayer perceptron, support vector machine can be improved by adding the data preprocessing blocks: orthogonal feature selection (Gram-Schmidt method) and nonlinear principal component analysis. We present the case study of various structures of deep learning systems (scenarios).