Pentru informaţii detaliate despre celelalte programe cofinantate de Uniunea Europeană, vă invităm să vizitaţi www.fonduri-ue.ro
DeepRiemann

# Beneficiary


Romanian Institute of Science and Technology - RIST


# Grant


The DeepRiemann project - Riemannian Optimization Methods for Deep Learning - is supported by the grant POC 2014-2020 - ANCSI Competitiveness Operational Programme 2014-2020. Priority Axis 1, Investment Priority P I 1a: Improving research and innovation infrastructures and capacities in order to develop excellence in RD&I and also promoting competence centres, particularly those of European interest, Action 1.1.4: Attracting high-level personnel from abroad in order to enhance the RD capacity, Programme code: POC-A1-A1.1.4-E-2015.

The total value of the project amounts to 8.689.500 lei, of which 7.276.617 comes from the European Regional Development Fund.

The project started on September 2016, and it will be completed on September 2020.


# Objectives


The DeepRiemann project aims at the design and analysis of novel training algorithms for Neural Networks in Deep Learning, by applying notions of Riemannian optimization and differential geometry. The task of the training a Neural Network is studied by employing tools from Optimization over Manifolds and Information Geometry, by casting the learning process to an optimization problem defined over a statistical manifold, i.e., a set of probability distributions. The project is highly interdisciplinary, with competences spanning from Machine Learning to Optimization, Deep Learning, Statistics, and Differential Geometry. The objectives of the project are multiple and include both theoretical and applied research, together with industrial activities oriented to transfer knowledge, from the institute to a startup or spin-off of the project.

More in details, the scientific objectives of the DeepRiemann project are:

  1. The definition of a probabilistic and geometric framework for the study of deep neural networks aimed at a better understanding of the working mechanisms behind the success of deep learning
  2. A complete characterization of the second-order Riemannian and affine geometries of a statistical model, aimed at the study of second-order optimization methods over statistical manifolds
  3. The design and implementation of new first and second-order methods for the optimization of functions defined over a statistical model, and in particular in the case of the training of neural networks
  4. An empirical evaluation of the algorithms proposed for the training of deep neural networks over standard datasets and in industrial competitions, aimed at showing the effectiveness of our methods with respect to the state of the art, in the fields of image and video, text document and audio document analysis
  5. A feasibility study, comprising a market analysis and the identification of novel potential markets for successful applications of deep learning
  6. The implementation of specific innovative demo applications based on deep learning technologies, and in particular on the algorithms developed during the project


# Members


Luigi Malagò project director
Cristian Daniel Alecsa postdoc researcher
Goffredo Chirco postdoc researcher
Deepika postdoc researcher
Hector Javier Hortua Orjuela postdoc researcher
Riccardo Volpi postdoc researcher
Alina Enescu research assistant and PhD student at Babeş-Bolyai University of Cluj-Napoca
Petru Hlihor phd student, co-supervised by Dr. Luigi Malagò (RIST) and Prof. Dr. Nihat Ay (MPI-MIS)
Csongor-Huba Várady phd student, co-supervised by Dr. Luigi Malagò (RIST) and Prof. Dr. Nihat Ay (MPI-MIS)
Alexandra Albu research assistant
Robert Colt research assistant
Delia Dumitru research assistant
Uddhipan Thakur research assistant
Sergiu Dahon busness developer
Larisa Calo project assistant


# Former members


Dimitri Marinelli postdoc researcher
Sabin Roman postdoc researcher
Septimia Sarbu postdoc researcher
Andrei Ciuparu research assistant
Titus Nicolae research assistant
Alexandra Pește research assistant
Raluca Drozan project assistant


# Submitted papers


  • Riccardo Volpi and L. Malagò.
    Natural Alpha Embeddings.

  • Hector J. Hortua, Riccardo Volpi, Dimitri Marinelli, and L. Malagò.
    Parameters Estimation for the Cosmic Microwave Background with Bayesian Neural Networks.

  • # Publications in journals


  • L. Malagò, L. Montrucchio, and G. Pistone.
    Wasserstein Riemannian Geometry of Positive Definite Matrices.
    In Information Geometry, Volume 1, Issue 2, pp 137-179, 2018

  • # Conference proceedings


  • Alexandra-Ioana Albu, Alina Enescu, and L. Malagò.
    Tumour Detection in Brain MRIs by Computing Dissimilarities in the Latent Space of a Variational AutoEncoder.
    In Proceedings of the Northern Lights Deep Learning Workshop, Septentrio Academic Publishing, 2020.

  • Petru Hlihor, Riccardo Volpi, and L. Malagò.
    Evaluating the Robustness of Defense Mechanisms based on AutoEncoder Reconstructions against Carlini-Wagner Adversarial Attacks.
    In Proceedings of the Northern Lights Deep Learning Workshop, Septentrio Academic Publishing, 2020.

  • # Workshop papers


  • S. Sârbu, R. Volpi, A. Pește, and L. Malagò.
    Learning in Variational Autoencoders with Kullback-Leibler and Renyi Integral Bounds.
    In ICML 2018 Workshop on Theoretical Foundations and Applications of Deep Generative Models, Stockholm, Sweden, 14-15 July 2018.

  • A. Pește, L. Malagò, and S. Sârbu.
    An Explanatory Analysis of the Geometry of Latent Variables Learned by Variational Auto-Encoders.
    In NIPS 2017 Workshop on Bayesian Deep Learning, Long Beach, US, 9 December 2017.

  • A. Pește and L. Malagò.
    Towards the Use of Gaussian Graphical Models in Variational Autoencoders.
    In ICML 2017 Workshop on Implicit Models, Sydney, Australia, 10 August 2017.

  • L. Malagò and D. Marinelli.
    Synthetic Generation of Local Minima and Saddle Points for Neural Networks.
    In ICML 2017 Workskop on Principled Approaches to Deep Learning, Sydney, Australia, 10 August 2017.


  • # Open positions


  • 1 postdoc positions on Deep Learning and Machine Learning, starting early 2020


  • # How to reach us


    The office is located in the city center of Cluj, only 9km away from the Cluj International Airport (Aeroportul International Avram Iancu Cluj). The simplest way to reach our office from the airport is to take a Taxi (around 25 lei). Alternatively, you can take the buses n. 5 or 8 in front of the airport, both stop at less than 10 minutes by walk from the office (ticket 2 lei). More information about the location of the bus stops can be found using Google Maps.






    Conţinutul acestui material nu reprezinta în mod obligatoriu pozitia oficiala a Uniunii Europene sau a Guvernului României

    (Last update 27 January 2020)