Presentations (Vanessa Haykal &  Taibou Birgui Sekou)

* Vanessa Haykal : 10h00

Title: A Combination of Variational Mode Decomposition with Neural
Networks on Household Electricity Consumption Forecast

Abstract: Recently, there has been a significant emphasis on the
forecasting of the electricity demand due to the increase in the power
consumption. This paper presents the computational modeling of
electricity consumption based on Neural Network (NN) training
algorithms. The noise in signals, which are caused by various external
factors, often corrupt demand series and influence consequently on the
model performance. For a more accurate electricity demand forecasting,
we propose a novel approach that combines a NN MLP (multilayer
perceptron) with VMD (variational mode decomposition)-based signal
filtering. By solving the constrained variational problem, the VMD
assumes that the signal is composed of a given number of modes where
each one is regarded as an amplitude-modulated and frequency-modulated
signal with its center frequency.

Using the daily electricity demand series, this paper demonstrates
that the proposed VMD-NN model greatly improves the forecasting error
comparing to existing stationary stochastic process such as the
autoregressive moving average (ARMA) model.

* Taibou Birgui Sekou

Title: Discriminative Dictionary Learning – Application to retinal
blood vessel segmentation

Abstract: Dictionary learning aims at learning a set of basis
functions that can best reconstruct a given family of signals in a
parsimonious manner. Since the pioneer work of Olshausen and Field in
the late 90’s, dictionary learning has now become state of the art
methodology on tasks such as face recognition, image denoising, image
up-scaling (a.k.a super resolution), to cite a few. In this work, we
compare some dictionary learning based methods with deep learning
based ones on retinal blood vessel segmentation. The results when
learning a dictionary reach recent deep learning based works and
sometimes outperform them. Dictionary learning being a representation
learning technique, we point out to future and ongoing work should dig
more in this direction and test on other modalities (MRI, CT scans,
etc.).