BrainMap

Mr. Bogdan Dumitrescu

Dr.ing.

- UNIVERSITATEA NAŢIONALĂ DE ŞTIINŢĂ ŞI TEHNOLOGIE POLITEHNICA BUCUREŞTI

Researcher | Teaching staff

Bogdan Dumitrescu was born in Bucharest, Romania, in 1962. He received the M.S. and Ph.D. degrees in 1987 and 1993, respectively, from University Politehnica of Bucharest, Romania. He is now a Professor with the Department of Automatic Control and Computers, National University of Science and Technology Politehnica Bucharest, where he works since 1990. He held several visiting research positions at Tampere University of Technology, Finland, in particular that of FiDiPro fellow (2010-2013). He was Associate Editor (2008-2012) and Area Editor (2010-2014) at IEEE Transactions on Signal Processing. He authored the book "Positive trigonometric polynomials and signal processing applications" and coauthored "Dictionary learning algorithms and applications". He is the author of 70+ journal papers, 100+ conference papers and 3 patents. His scientific interests are in optimization, numerical methods, and their applications to signal processing, especially using sparse representations.

Expertise & keywords

Asymmetric Dictionary Learning Call name: P 4 - Proiecte de cercetare exploratorie - PCE-2021
PN-III-P4-PCE-2021-0154 2022 - 2024

Role in this project:

Coordinating institution: UNIVERSITATEA POLITEHNICA DIN BUCURESTI

Project partners: UNIVERSITATEA NAŢIONALĂ DE ŞTIINŢĂ ŞI TEHNOLOGIE POLITEHNICA BUCUREŞTI (RO)

Affiliation:

Project website: http://asydil.upb.ro/

Abstract:

Dictionary learning (DL) is a technique for adapting sparse representations to training data, with many applications in signal and image processing and also in machine learning tasks like classification. Although very versatile, DL essentially builds linear representations, a fact which limits performance in certain applications. Some steps towards nonlinearity have been made, notably kernel DL, but the topic is largely open.

We propose a structural extension to DL: to replace the dictionary atoms, which are vectors, with sets, for example cones around a central vector. A single atom is chosen from an atom-set for sparse representation. This extensions significantly increases non-linearity and poses interesting optimization problems for the computation of the best sparse representation and for DL itself. It can be applied to standard DL as well as to kernel or other DL forms.

The new structure is called Asymmetric DL, due to the lack of reconstruction capability in the absence of the represented signal. However, many applications are still possible, starting with denoising and classification. Our focus will be on anomaly detection, with special interest to graph data coming from bank transactions, on which we have access to real data.

A goal of the project is to make MATLAB and Python libraries containing all the new algorithms.

The research team at University Politehnica of Bucharest is made of a professor, a postdoc, two PhD students and a master student.

Efficient Learning and Optimization Tools for Hyperspectral Imaging Systems Call name: EEA Grants - Proiecte Colaborative de Cercetare
RO-NO-2019-0184 2020 - 2023

Role in this project:

Coordinating institution: UNIVERSITATEA POLITEHNICA DIN BUCURESTI

Project partners: UNIVERSITATEA NAŢIONALĂ DE ŞTIINŢĂ ŞI TEHNOLOGIE POLITEHNICA BUCUREŞTI (RO); Norwegian University of Science and Technology (NO); UNIVERSITATEA BUCURESTI (RO); SPITALUL CLINIC COLTEA (RO)

Affiliation:

Project website: https://elohyp.wordpress.com

Abstract:

With the increasing amounts of data around us, many modern technologies designed for hyperspectral image processing, computer vision and medical imagining are now routinely using Artificial Intelligence (AI) and Big Data techniques. Hyperspectral imaging collects and processes information from the electromagnetic spectrum using a subset of targeted wavelengths at chosen location that span beyond the usual RGB spectrum. Hyperspectral data are becoming a valuable tool for monitoring the Earth’s surface or human body and are used in a wide array of applications: agriculture, health, environment, mineralogy, surveillance, physics, astronomy and chemical imaging. In the last decades, a large number of methods were proposed to deal with image processing problems. However, most these methods have been designed for application to color and/or grayscale images; therefore, they have limited success when applied to hyperspectral images. This is partially owing to large hyperspectral datasets being difficult to collect, process and analyze, and also to the heavy computational load associated with images captured using many spectral bands.

ELO-Hyp will find practical answers to these scientific challenges. We will push the frontiers of AI and Big Data and take important steps towards making the hyperspectral technologies of tomorrow possible, with direct impact in various parts of environment, health and industry sectors. To address these difficulties, we propose novel learning and optimization techniques for hyperspectral imaging systems. Our first novelty consists in defining appropriate models (loss functions), which include additional data representations, relational information about data (e.g. taking into account not only spectral information, but also spatial information) or ways of removing the influence of noise from predictor performance, boosting performance even when dealing with small datasets. However, high-quality models based on these new features would require hard Big Data (possibly nonconvex) optimization formulations, which further need fast and scalable algorithms with proper convergence guarantees. Therefore, our second novelty consists of developing fast optimization algorithms (e.g. higher order, projection or stochastic splitting methods) having low complexity per iteration (e.g. by combining with coordinate descent framework) and scalability (e.g. using parallel architectures such as GPUs, FPGAs). In conclusion, we aim to create, analyze and implement efficient learning and optimization algorithms for hyperspectral imaging models with applications to ocean monitoring and medical imaging. We will implement the algorithms, benchmark them using real datasets, ensure the algorithms’ interoperability, and produce free software.

Structured Linear Algebra for Sustainable Machine Learning Call name: P 1 - SP 1.1 - Proiecte de cercetare pentru stimularea tinerelor echipe independente
PN-III-P1-1.1-TE-2019-1843 2020 - 2022

Role in this project:

Coordinating institution: UNIVERSITATEA POLITEHNICA DIN BUCURESTI

Project partners: UNIVERSITATEA POLITEHNICA DIN BUCURESTI (RO)

Affiliation: UNIVERSITATEA POLITEHNICA DIN BUCURESTI (RO)

Project website: https://cs.unibuc.ro/~crusu/slam.html

Abstract:

This project develops new numerical methods based on structured linear algebra with applications to efficient algorithms for signal processing, machine learning, and artificial intelligence. State-of-the-art machine learning techniques are revolutionizing many fields of science (computer science, biology, economics, business, to name a few) and the society at large by taking over increasingly complex tasks from human operators. While it is true that these technological breakthroughs have immense benefits and promise to revolutionize the global economy, unfortunately, they also exhibit at least one major, potentially crippling, drawback: high costs of operating the state-of-the-art machine learning methods on specialized hardware infrastructure and with high costs in terms in energy consumption (direct monetary costs and indirect costs in the form of damage to the environment). We are quickly reaching a situation where only a few large companies or well-funded research groups can afford the costs of running large-scale state-of-the-art machine learning methods. Furthermore, even in these cases, the damage done to the environment due to CO2 emissions is unacceptable. Therefore, lowering the costs associated with modern machine learning methods is essential to ensure their widespread adoption and their longterm sustainability. As machine learning methods heavily rely on numerical linear algebra (and mostly matrix calculations), in this project, we propose new numerical methods based on new structured matrices to reduce the storage and computational complexity of training and running machine learning models. We expect our work to have a direct impact on the latest machine learning techniques, such as optimization techniques, dictionary learning for sparse representations, kernel methods, and neural networks.

Stochastic Optimization Algorithms for Anomaly Detection Call name: P 1 - SP 1.1 - Proiecte de cercetare Postdoctorală
PN-III-P1-1.1-PD-2019-1123 2020 - 2022

Role in this project: Key expert

Coordinating institution: UNIVERSITATEA BUCURESTI

Project partners: UNIVERSITATEA BUCURESTI (RO)

Affiliation: UNIVERSITATEA BUCURESTI (RO)

Project website: https://sites.google.com/site/andreipatrascuro/grants?authuser=0

Abstract:

With the ever increasing amounts of data around us, many modern technologies designed for anomaly detection, computer vision or pattern recognition are now routinely using machine learning and data mining techniques. Anomaly detection is the problem of distinguishing between normal data samples with well designed patterns or signatures and those that do not conform to the expected profiles. Unsupervised classification and decomposition models shows more and more success on fraud or intrusion detection applications as it is emphasized by valuable literature sectors. Promising research directions in this branch move towards eliminating the influence of outliers from predictor performance, or towards inclusion of additional descriptions of data or relational information about data that could boost the model performance. However, since the high quality models based on these new features would require hard stochastic optimization formulations, in StOpAnomaly project we aim to create, analyze and implement stochastic splitting algorithms for composite optimization problems with focus on robust anomaly detection models based on decomposition and one-class classification, empowered with background information. On short, the outlier influence will be minimized thorugh finding specific robustness-promoting loss functions and the background information synthesis will be approached by Privileged Information or Graph Embedding.

Graphomaly - software package for anomaly detection in graphs modeling financial transactions Call name: P 2 - SP 2.1 - Proiect experimental - demonstrativ
PN-III-P2-2.1-PED-2019-3248 2020 - 2022

Role in this project: Project coordinator

Coordinating institution: UNIVERSITATEA POLITEHNICA DIN BUCURESTI

Project partners: UNIVERSITATEA POLITEHNICA DIN BUCURESTI (RO); UNIVERSITATEA BUCURESTI (RO); TREMEND SOFTWARE CONSULTING SRL (RO)

Affiliation: UNIVERSITATEA POLITEHNICA DIN BUCURESTI (RO)

Project website: http://graphomaly.upb.ro

Abstract:

The proposed project, called Graphomaly, aims to create a Python software package for anomaly detection in graphs that model financial transactions, with the purpose of discovering fraudulent behavior like money laundering, illegal networks, tax evasion, scams, etc. Such a toolbox is necessary in banks, where fraud detection departments still use mostly human experts.

The main tool that we propose is dictionary learning for sparse representations, which will be used to model sub-graphs derived from the full transactions graph through community detection. Other machine learning tools will be used for comparison, together with a set of data processing tools that are customary for dimensionality reduction.

There are two main working scenarios. In one, fraud patterns are known, but their shape can vary in size and also can be affected by other activities. In the other, unsupervised learning is used for the detection of anomalies, possibly of new types, that may be related to frauds.

The implemented methods will be able to process large graphs. Online and distributed forms of the algorithms will be derived, such that reaction time is decreased and thus frauds can be discovered in their incipient stages.

The consortium is made of two universities and a software firm and has the support of a bank that will provide relevant transactions data and will directly validate some of the results. The team members have relevant expertise in dictionary learning and related techniques, software architecture, data management and processing.

LOcal prediction Watermarking – Second Generation Call name: P 1 - SP 1.1 - Proiecte de cercetare Postdoctorală
PN-III-P1-1.1-PD-2016-1666 2018 - 2020

Role in this project: Key expert

Coordinating institution: UNIVERSITATEA "VALAHIA" TARGOVISTE

Project partners: UNIVERSITATEA "VALAHIA" TARGOVISTE (RO)

Affiliation: UNIVERSITATEA "VALAHIA" TARGOVISTE (RO)

Project website: http://low2g.valahia.ro

Abstract:

Watermarking is the imperceptible embedding of data into digital host signals like image, video, audio, text, etc. Its applications are in copyrighting, authentication and annotation. While classical watermarking introduces permanent distortions, reversible watermarking (RW) not only extracts the embedded data, but also recovers the original host signal/image without any distortion. The most efficient RW schemes for digital images embed data into the prediction error. This makes prediction improvement the major research direction in RW.

The best prediction scheme for RW is our local prediction approach (published in IEEE Transactions on Image Processing, 2014) that computes, for each pixel, its own least mean square predictor in a local block. The same predictor is recovered at detection without any additional information.

The main objectives of the LOW-2G projects are to:

1) Develop novel prediction schemes. Instead of fixed prediction parameters as considered in the original local prediction scheme, we shall investigate the use of the adaptive parameters for the learning block sizing, the adaptive changing of the size and/or shape of the prediction context and different optimization techniques for local predictor computation (least mean squares (LMS), iteratively reweighted least squares (IRLS), etc.).

2) Develop a new generation of RW schemes based on local prediction and investigate their applications. The novel prediction schemes will be used to develop a new generation of RW schemes of lower distortion than the state-of-the-art ones. Gains of 2-4 dB are expected. Fast algorithms and fast implementations of the proposed schemes will be also investigated. New applications taking advantage of the lower distortion introduced by the new LOW-2G schemes will be developed in the framework of the project.

Sparse representations in signal processing Call name: Exploratory Research Projects - PCE-2011 call
PN-II-ID-PCE-2011-3-0400 2011 - 2016

Role in this project: Project coordinator

Coordinating institution: Universitatea Politehnica Bucuresti

Project partners: Universitatea Politehnica Bucuresti (RO)

Affiliation: Universitatea Politehnica Bucuresti (RO)

Project website: http://www.schur.pub.ro/Idei2011.htm

Abstract:

A model is sparse if it has only few nonzero parameters. The latest decade has seen significant research effort in analyzing the properties of sparse representations, in developing algorithms for their computation and in using them in signal processing applications. However, the field is still open to innovation, both in theory and applications.

In this project, we aim to investigate the following problems.

1. Sparse filters design. The purpose is to develop algorithms for obtaining 1D and 2D sparse FIR filters that are better than those currently available and to extend the results to filter banks.

2. Sparse total least squares for solving linear systems: theoretical properties, greedy algorithms and the relation with least squares.

3. Distributed sparse filtering. We plan to find algorithms for adaptive filtering, like recursive least squares (RLS), that are fit to compute sparse solutions in a distributed environment, with access only to local data and restrained communication.

4. Classification with sparse predictors. We want to find or adapt dictionary optimization algorithms and to obtain dictionaries that allow discrimination between several categories of audio signals. The main employed audio feature is the sparse predictor, which gives good local information, possibly combined with global features.

The research team is composed of a professor, a postdoc and a PhD student, with experience in signal processing, convex optimization and numerical methods.

FILE DESCRIPTION	DOCUMENT
List of research grants as project coordinator
List of research grants as partner team leader
List of research grants as project coordinator or partner team leader
Significant R&D projects for enterprises, as project manager
R&D activities in enterprises
Peer-review activity for international programs/projects