Prof. Jianhua Zhang, Oslo Metropolitan University, Norway

Jianhua Zhang is has been Professor at Department of Computer Science, Oslo Metropolitan University, Norway, since 2018. From 2007-2017 he was Professor with School of Information Science and Engineering, East China University of Science and Technology, Shanghai, China. From 2017 to 2018 he was Scientific Director at Vekia (a machine learning driven IT company), Lille, France.
Dr Zhang received his PhD in electrical engineering and computer science from Ruhr University Bochum, Germany, in 2005 and did postdoctoral research at Intelligent Systems Research Lab, University of Sheffield, UK, from 2005 to 2006. He was a Guest Scientist at TU Dresden, Germany, from 2002 to 2003 and Visiting Professor at TU Berlin, Germany between 2008 and 2015 and the University of Catania, Italy in 2024.
Dr Zhang has worked in the fields of AI, control systems, and signal processing since mid-1990s. His current research interests include computational intelligence, machine learning, intelligent systems and control, biomedical signal processing, and neurocomputing. So far he has published four books, 11 book chapters, and around 200 peer-reviewed journal and conference papers in those areas.
Dr Zhang served as Chair of IFAC (International Federation of Automatic Control) Technical Committee on Human-Machine Systems (2017-2023) and Vice Chair of IEEE Norway Section (2019- 2023). He currently serves as Vice Chair of IFAC Technical Committee on Human-Machine Systems (2023-) and Vice Chair of IEEE CIS (Computational Intelligence Society) Norway Chapter (2019-). He is on editorial board of four international journals, including Frontiers in Neuroscience, Cognitive Neurodynamics (Springer), and Cognition, Technology & Work (Springer). In addition, he was invited to serve as keynote speaker or chair for a number of international conferences.
Dr Zhang was listed in Stanford/Elsevier's World Top 2% Scientists Rankings in 2023 and 2024.

Speech Title: Stock Price Forecasting by Means of Transformer-based Ensemble Learning

Abstract: In this talk, for the stock price forecasting problem we compare the performance of several models, including traditional time series analysis model - ARIMA and four machine learning (ML) models (Linear Regression, Long Short-Term Memory (LSTM) network, Prophet, and Transformers). Ensemble learning is proposed to reduce the prediction biases and variances of those individual models. Furthermore, in order to handle the complexity and volatility of real-world stock markets, three different hyperparameters (such as learning rate, number of layers in the network model, batch size, etc.) tuning approaches (grid search, random search, and Bayesian optimization) are compared in terms of prediction accuracy and computational cost. The real stock data analysis results showed that ensemble learning method can improve accuracy and reliability of stock time series forecasting and that the transformer model stacked with linear regression achieved the best prediction performance. The results obtained may provide insights into stock closing price dynamics modeling, stock investment decision, and portfolio management.

Prof. Sami Brandt, IT-University of Copenhagen, Denmark

Prof. Sami S. Brandt works as full professor at IT University of Copenhagen and he is the head of the audio–visual computing research group at ITU. He is also currently the project coordinator of the EU funded Horizon Europe project XTREME. Previously, he worked as associate professor at ITU and University of Copenhagen, and senior mathematical software developer at 3Shape, Denmark; He has also been nominated as adjunct professor in Machine Vision Group at University of Oulu, Finland. He additionally has project leading experience from Synarc Imaging Technologies A/S, University of Oulu and Helsinki University of Technology, Finland, and he has worked as a research scientist in Malmö University, Sweden, and in Instrumentarium Corporation Imaging Division. He finished his doctoral degree in 2002 in Helsinki University of Technology about the geometric branch of computer vision appled to electron tomography. His research interests are Applied mathematics, Bayesian inverse problems, 3D imaging, mathematical modelling, statistics, geometric computer vision, image processing, image registration, acoustics and audio signal processing, biomedical imaging, and machine learning.

Speech Title: On the non-rigid structure-from-motion problem: from independent subspace analysis, degenerate basis shapes, and tensor-based factorisation to generative adversarial networks

Abstract: This talk provides an overview of our work on the non-rigid structure and motion problem with the application of the modeling and analysis of human faces. We start by presenting the classic formulation of the problem where the goal is to estimate the non-rigid affine structure and motion from 2D point correspondences and note its known difficulties and approaches taken to tackle them. Thereafter we show how independent subspace analysis can help to achieve a solution where no prior formation, apart from the assumption of statistical independence of the basis shapes, nor camera calibration information is required. Thereafter we develop another solution to the problem, by modifying the common assumption that the non-rigid shape is a linear combination of basis shapes, by adding an additional constraint, that the basis shapes should be degenerate. By this assumption, it is then possible to derive a solution that avoids the central problems of the classic problem setting. We likewise show, how tensor based modelling of faces and non-rigid structure-motion-problem can be unified into single tensor-based modelling problem. In the last part of the talk, we show how our non-rigid structure-from-motion approaches can be extended to generative models such as StyleGAN model to achieve factorization of latent manifolds into camera geometry, pose, and non-rigid structure that opens the way of photorealistic modelling, analysis and editing of human faces and the underlying geometry via the trained generator. The future directions are also discussed.