Our Company

Libra MLI, is a research intensive SME well qualified to bring cutting-edge Machine Learning theory into practice. It comprises personnel, which is highly experienced in Machine Learning state-of-the-art research for several years, having in its portfolio a long track record in top scientific publications and presence in the major workshops and conferences in the field.

LIBRA specializes in Machine Learning:

  • for online/streaming data or for situations where the data statistics dynamically changes with time (online learning)
  • for data which have some entries wrong, mixed up, corrupted or comprise missing values (robust learning)
  • for data which are spatially distributed in several locations and their transmission through the network is not allowed due to security and/or privacy reasons (distributed learning)
  • for applications where low computational complexity and reduced storage requirements are critical, e.g. when Big Data is involved

Applications

Machine learning technologies exhibits a large generalization potential being suitable, after limited modifications, for a large number of diverse applications. Indicative examples are:

  • Multimodal data analysis, e.g. FMRI data analysis, Hyperspectral image analysis, DNA sequencing, learning analytics
  • Recommender systems, e.g. audio recommendations, product recommendations, automatic data completion
  • Pattern recognition and classification, e.g. audio classification, handwritten digital recognition, object recognition
  • Anomaly detection, i.e. detection of signs and patterns which indicate deviations from normal or expected conditions, e.g. behavioural change detection, epilepsy onset detection, spam and fraud detection
  • Predictive analytics, e.g., trends prediction, clinical decision support systems, financial and behavioural analytics

Services

Currently, LIBRA offers the following services:

  • Consultancy, covering everything required to implement Machine Learning tailored to specific data and conditions, such as suggestions about Machine Learning approach and platform, assessment of pattern analysis, predictive analytics methodology requirements and scalability planning. Moreover, LIBRA offers optimization of commercially available Machine Learning Methods and platforms in order to best suit the data and targets at hand.
  • Data analysis services based on the LIBRA algorithmic suite, including custom ML modeling and algorithm design, implementation architecture and integration, analytics engine deployment and operation.
  • Serving as partner in ambitious research projects where Machine Learning can play a distinct role in a variety of applications

The Core Team


Ioannis (Yannis) Kopsinis, Phd

LIBRA director, co-founder. He received his PhD from the Dept. of Informatics and Telecommunications, Univ. of Athens in 2003. Since then he has gained a number of prestigious personal research grants. Among them, a Marie Curie IEF fellowship, Univ. of Athens, Greece and a Ramón y Cajal Fellowship, University of Granada, Spain. Moreover, he has worked for more than 5 years as a senior research fellow in the School of Engineering and Electronics, the University of Edinburgh, UK. He has published more than 50 papers in technical journals and conferences and he has co-authored 3 book chapters. His current research interests lie in the areas of Online Machine Learning, Constrained optimization and robust predictive analytics.

Pantelis Bouboulis, Phd

LIBRA Scientific Manager. He graduated from the mathematics department of the National and Kapodistrian University of Athens, Greece, at 1999. He received his M.Sc. and Ph.D. degrees from the department of Informatics and Telecommunications of the same university in 2002 and 2006 respectively. From 2007 till 2008, he served as an Assistant Professor in the same department. From 2013 to 2015 he worked as a research assistant for the ASSURANCE project (funded by Greece and the EU under the Aristeia framework). From 2008 to 2015, he also taught mathematics at several model-experimental Greek high schools (organizing STEM related science clubs). He has co-authored 20 articles in various international scientific journals, he has received the ICPR Best paper award (2010) and two scholarships. His current research interests lie in the areas of Machine Learning, Fractals, Signal and Image Processing. Since 2010 he serves as an Associate Editor of the IEEE Transactions on Neural Networks. He is a member of the ΙΕΕΕ and the Computational Intelligence Society.

Robust Learning

The majority of Machine Learning algorithms, require data to be well conditioned and without misleading entries. However, this is a luxury in most of the applications especially in those related with Big Data environments. Data sets are very likely to be corrupted, in a smaller or a larger extent, either unintentionally or intentionally; Humans make sporadic mistakes, so hand-filled data are likely to have erroneous inputs, sensors can be corrupted leading to wrong outputs or no output at all, malicious attacks can lead to unwilling data modifications, old archives, e.g. medical records, are likely to be half-complete having many missing entries, misleading entries can be given in test/examination data, e.g. people participating in surveys might answer randomly or intentionally erroneously, or people participating in exams might cheat. Corrupted data is the norm rather than the exception and such data entries appear as outliers, i.e. values which cannot be explained from the overall data statistics. Machine Learning techniques which do not deal with outliers, meaning that they take all data for granted, are likely to lead to wrong analysis results and diminishing performance.

LIBRA approach aims at handling outliers explicitly, managing to detect and isolate them or even correct them, if this is feasible. As a result, the LIBRA Machine Learning toolbox does not heavily depend on the performance and accuracy of any pre-processing or preconditioning stage before the learning function.

Big Data

Data volumes are exploding; more data has been created in the past two years than in the entire previous human history! The Big Data era we are now entering is characterized by massive data volumes, large variety of data types, excessive velocity as well as a large uncertainty regarding their validity and correctness. In any case, for Machine Learning, large data volumes is a luxury as long as one can exploit their full potential by being able to process them and analyse them in order to arrive to better decisions, improved performance, novel applications and deeper insights.

Because of the massive data sizes, their increased complexity (since new data types emerge), their veracity, speed as well as their involvement in the Internet of Things paradigm, many of the traditional techniques run into limitations.

LIBRA is developing and constantly enriching a Machine Learning algorithmic suite, which can cope with most of the Big Data challenges, exploiting their potential for online/streaming operation, their robustness as well as their scalability through distributed decentralised learning.

Distributed Learning

In principle, distributed Learning refers to all those Machine learning approaches and variants which aim at distributing the overall computational load and/or data load of a Machine Learning task into several processing/storage units. Distributed approaches for Machine Learning become essential in the Big Data era. Most of the Distributed Machine Learning solutions today are centralized meaning that a central node has access to all data and governs/organizes the whole process.

LIBRA, on the contrary, specializes on fully distributed/decentralized solutions which offer certain advantages:

  • The existence of a fusion center is avoided and one solely relies on in-network processing in ad-hoc topologies. This leads to increased reliability and robustness of the network system, because it is not affected by possible fusion center failures.
  • Data is stored locally without the necessity of some nodes having access to all data. This approach enhances privacy since the processing of data is performed locally, avoiding the need for sensitive information exchange.

Moreover, LIBRA supports Online Distributed Learning where the data might be streaming in spatially distributed nodes.

Online Learning

The majority of Machine Learning technologies perform batch processing of the full data set, meaning that in order for the processing to begin, the full amount of data needs to be collected and stored. Online learning is different in that the data, instead of getting stored, is processed sequentially, “on the fly”, as soon as it becomes available.

With online learning, the system performance is improving with time as more and more data is processed. Moreover, online Machine Learning is naturally adaptive having the ability to continuously adjust to the data dynamics and characteristics and it can effectively handle cases where the data essence and latent patterns are changing with time.

Online learning appears to be an important ingredient in current and future Big Data applications.

Machine Learning

In humans, learning is the act or the process of acquiring new knowledge, skills and the ability to infer, which is accomplished using information, i.e. processed and organized data, available in the environment.

Machine Learning refers to technologies and mathematical methods which empower algorithms running in computers or other devices with the capability to learn from data and discover insights, unveil hidden structures, regularity patterns, correlations as well as possible anomalies. Such invaluable information is in turn used to make predictions, make inference about data unseen so far, detect information-bearing events, classify/organize similar data, mobilize decision support systems and analyse data into their fundamental latent constituents reflecting their generation mechanism and nature.

Machine Learning technologies are now key constituents of many current and future applications in a great variety of disciplines.

Machine Learning

In humans, learning is the act or the process of acquiring new knowledge, skills and the ability to infer, which is accomplished using information, i.e. processed and organized data, available in the environment.

Machine Learning refers to technologies and mathematical methods which empower algorithms running in computers or other devices with the capability to learn from data and discover insights, unveil hidden structures, regularity patterns, correlations as well as possible anomalies. Such invaluable information is in turn used to make predictions, make inference about data unseen so far, detect information-bearing events, classify/organize similar data, mobilize decision support systems and analyse data into their fundamental latent constituents reflecting their generation mechanism and nature.

Machine Learning technologies are now key constituents of many current and future applications in a great variety of disciplines.

Robust Learning

The majority of Machine Learning algorithms, require data to be well conditioned and without misleading entries. However, this is a luxury in most of the applications especially in those related with Big Data environments. Data sets are very likely to be corrupted, in a smaller or a larger extent, either unintentionally or intentionally; Humans make sporadic mistakes, so hand-filled data are likely to have erroneous inputs, sensors can be corrupted leading to wrong outputs or no output at all, malicious attacks can lead to unwilling data modifications, old archives, e.g. medical records, are likely to be half-complete having many missing entries, misleading entries can be given in test/examination data, e.g. people participating in surveys might answer randomly or intentionally erroneously, or people participating in exams might cheat. Corrupted data is the norm rather than the exception and such data entries appear as outliers, i.e. values which cannot be explained from the overall data statistics. Machine Learning techniques which do not deal with outliers, meaning that they take all data for granted, are likely to lead to wrong analysis results and diminishing performance.

LIBRA approach aims at handling outliers explicitly, managing to detect and isolate them or even correct them, if this is feasible. As a result, the LIBRA Machine Learning toolbox does not heavily depend on the performance and accuracy of any pre-processing or preconditioning stage before the learning function.