Research topics
Brain
Modeling and Simulations: Olfactory Bulb: The understanding of the
nonlinear dynamics of olfactory bulb (OB) is essential to the modeling of brain
and nervous system. Odor threshold, odor identification, detection and
recognition are the basic measures of the medical studies detection and
diagnosis of neuro-degenerative diseases as Alzheimer’s, Parkinson’s disease
and schizophrenia.
On the base of
our study of the OB activities and the analysis of the conditions governing
neural oscillations and the nature of odor-receptor interactions, we proposed
and developed mathematical models and simulations addressing the questions of
how the brain recognizes odors, how it works in a noisy natural environment and
why synchronization is used for decoding brain circuits, which are still not
successfully solved. We simulated the
dynamic behavior of the olfactory system in order to understand the way in
which odors are represented and processed by the brain.
Radial Basis Function Neural Networks Based on Potential Functions: A new strategy of shape-adaptive radial basis functions based on potential functions and optimization procedure for positioning of the centers during the learning process is suggested. The novelty of the presented approach includes: 1) Automatic capacity correction of functions separated the cluster and automatic NN topology selection in the sense that both the optimal number and locations of the basis functions are automatically obtained during training. 2) Extracting of a small number of supporting patterns from the training data that are relevant to the classification what further improves the convergence.
Self-Organizing Maps:
Classical Self-Organizing Maps read the input values in random but sequential
order one by one and thus adjust the network structure over space: The network
will be built while reading bigger and bigger parts of the input. In contrast
to this approach, we present a Self-Organizing Map that processes the whole
input in parallel and organizes itself over time. The main reason for parallel
input processing lies in the fact that knowledge can be used to recognize parts
of patterns in the input space that have already been learned. This way,
networks can be developed that do not reorganize their structure from scratch
every time a new set of input vectors is presented but rather adjust their
internal architecture in accordance with previous mappings.
Neural Network Models for Independent Component
Analysis: ICA is a technique, which transforms a multivariate data vector
into a new one whose components are as statistically independent as possible.
We propose a multiplayer neural network structure based for feature extraction
based on ICA. We built a graphical user interface (GUI) in order to impose
different type of noise to the experimental images and extract the uncorrelated
components from natural scenes. The
analysis of the results shows that the proposed multilayered feedforward neural
network for image feature extraction using ICA significantly reduces the noise
by elimination of some small features which are in the original image.
Feature Selection Fuzzy Algorithms in
High Dimension Data Clustering: The
effectiveness of pattern clustering (recognition) is highly dependent on the
accurate identification of clusters shapes which can be influenced by considering
prototypes with a geometric structure or by using different distance measures
in the objective function. We developed
an extension of FCM algorithm by using Mahalanobis (MFCM) and parametric
Minkowski (&MFCM) dissimilarity distances which are suitable for processing
datasets with high dimensions and high number of clusters. The proposed
initialization and use of validation measures to determine the initial number
of clusters is very effective. We also integrate Sammon mapping and Silhouette
function for data visualization and analysis of the clustering results.
Algorithms for Handling Dataset Imbalance
in Neural Networks with Deep Learning: We focus on the topology
and training of neural network classifiers
with deep learning when using imbalance data and different ways of data
preprocessing, implementation of some newly suggested optimizers as well as
traditional one for Deep Learning (DL), different ways of learning rate
adaptation including the one based on cyclic increase and decrease during the
training and propose a taxonomy of modelling strategies for handling imbalanced
datasets.