Natural and High Performance
Computing Laboratory (NHPC Lab)
http://atm.neuro.pub.ro/radu_d/
Team:
Radu Dogaru, Ph.D., Full professor, PhD
advisor (computational intelligence, natural computing,
nonlinear dynamics)
Ioana Dogaru, Ph.D., Associate professor, (advanced
programming techniques and tools, reconfigurable computing
systems, natural computing)
2 teaching staff (PhD, listed above) + 7 (variable by the year) PhD students/researchers
1 office and research lab + 1 educational and research lab with up to 10 work-places.
Research
directions:
- Natural computing algorithms (cellular nonlinear networks, artificial neural networks and machine learning, fuzzy logic, swarm intelligence and naturally inspired optimization and learning algorithms, etc.) and their applications
- Nonlinear dynamics in complex networks and their applications: ciphering, prediction, models for various physical, biological and social processes.
- Embedded solutions with HP3 (high performance, portability and productivity) for implementing the above mentioned algorithms into actual computing platforms (PC, FPGA-based systems, GPU based systems with CUDA support, development of specialized circuits in cooperation with other groups)
Representative
results:
1. The RBF-M (recently renamed FSVC = Fast Support Vector Classifier) - more details follow
2.
The Simplicial
Neural
Network (based on it a smart visual
processor was recently (2014) developed: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6716069);
3.
The SORT
neural network (or a brief version here)
it is a simplified version of the Simplicial NN targeting
digital platforms.
4.
The modified
Naive-Bayes classifier (it
is actually a Naive Bayes classifier with tuning parameters,
thus allowing better performance). It also includes easy-to
implement triangular shape functions replacing the traditional
Gaussian.
5.
(most cited of our papers)
The Multi-nested neuron
(universal
CNN cell); Initially developed to provide a very compact
implementation of CNN cells, may be also used in any other
machine learning task. It is a very compact, hardware
oriented, alternative to polynomial neurons.
Some
relevant projects and downloads
- FSVC -
Efficient classifier for embedded systems (Fast
Support Vector Classifier) - freely available
on
MatlabCentral. Simillar performance to well known SVM
but with a much lower complexity and avoiding complicated
operators (exponential, etc.) - well suited for embedded
systems. Related papers related are included in the above link.
Faster version is accelerated two orders of magnitude
using .MEX files (runs under Octave 4.0.0). Available
on GitHub.
SFSVC: Another (even faster, 10-50 times more than SVM
and ELM on the same platform) was presented in a paper
accepted at the COMM 2016 conference); SFSVC is well suited for
embbeded systems, there is no need for updating synaptic
weights, while support vectors are simply selected from the
training dataset. A Python version optimized for speed
will be soon (july 2018) made available on GitHub
- ADBIOSONAR
- research project (2008-2011) financed by the Romanian
Research Founding Authority, concluded with a cellular automata based
software simulator for
virtual experiments using sound propagation (initially
conceived for robotics). Related papers here.
- AutoVoice- Database: A database with
vocal commands in noisy environment (inside an
automobile) made by Mihai Bucurica as part of his doctoral
studies. More details here
(including a paper reporting research using this database,
please the cite this paper if you find it useful for your
research).
April 2018 - "Some
fast and compact neural network solutions for artificial
intelligence applications"
Invited lecture given at the