Vous êtes ici : Accueil » Kiosque » Annonce

Identification

Identifiant: 
Mot de passe : 

Mot de passe oublié ?
Détails d'identification oubliés ?

Annonce

5 juin 2020

Generative Neural Networks for Learning Stochastic Systems applied to adaptive deep-brain stimulation -- Réseaux de neurones génératifs pour l'apprentissage des systèmes stochastiques appliqués à la stimulation adaptative du cerveau profond


Catégorie : Doctorant


Thématique / Domaine / Contexte
Basic AI and Data Science: statistical learning theory for high dimensional data
Specialized ML and AI: signal, Boltzmann machines, stochastic models, adaptive control, deep-brain stimulation
Application domain: Health and well-being, neurodegenerative disease
Key words
stochastic models, adaptive control, deep-brain stimulation, neural networks

Contact: Vincent Vigneron, Hichem Maaref
(vincent.vigneron,hichem.maaref)@ibisc.univ-evry.fr
Phone: +33 6 635 687 60

 

Description of the research problem
The Pakinson's Disease (PD) is one common neurological diseases for the elderly, and is attributed to the degeneration of dopaminergic neurons in the substantia nigra pars compacta (SNpc).
Although medicine are available for replacing the degenerated, dopaminergic connections, medication becomes less or even adversely effective as the disease progresses. A promising alternative is applying high-frequency (>100Hz) stimulation to the subthalamic nucleus (STN) through microelectrodes chronically implanted in the deep brain [1,2]. The deep-brain stimulation (DBS) has been approved for treating PD, and some studies suggest that the DBS could act as a transient lesion that blocks dysfunctional regions (e.g. the STN). However, themechanism of DBS remains not fully understood yet. The clinical application is limited to applying permanent, periodic stimulation, which has been found to induce several neuropsychiatric side effects [3].
Therefore, understanding the DBS mechanism is crucial for developing more effective stimulation techniques with reduced side effects. The effectiveness refers to coordinating the stimulation in a closed-loop (on-demand) and sophisticate manner. The stimulation is turned on only when the brain enters the diseased state, and the stimulation could be applied tomultiple sites with sophisticate waveforms to disrupt abnormal neural activities more quickly.
To achieve this goal, it is crucial for developing a computational model able to simulate the PD-pertinent neural network quantitatively with satisfactory fidelity.
Our team has established the PD rat model and developed a electronic circuit for recording/stimulating multiple regions of a rat brain simultaneously and chronically. In addition, several new nuclei pertinent to the PD have been identified. By analyzing the activities recorded from multiple brain regions of normal and PD rats before and after deep-brain stimulation (DBS), this thesis aims to (1) develop algorithms able to identify signatures that indicate reliably the transition of the rat brain from normal to diseased states and vice versa, and to (2) model how the DBS modulates the activities of the PD-pertinent neural network.
The first objective would underpin a better (closed-loop) control on the stimulation timing for treating PD. The second objective would lead to better understanding on the DBS mechanism, and would subsequently indicate more effective ways of treating PD by DBS.

Objective
The PhD thesis aims to develop an evolutionary algorithm capable of reproducing the main stochastic dynamical properties of a complex system (e.g. brain circuitry) automatically. As the evolutionary dynamic model does not consist of physical equations like conventional models (e.g. Kalman model), the observability and controllability properties of the evolutionary dynamic model will be investigated

Method
Based on the stochastic model called conditional Restricted Boltzmann Machine (CRBM) (due to Hinton works in the 90's), its ability to model multi-dimensional local field potentials recorded from therat brain will be first examined. Afterward, the on-line learning algorithm for the CRBM will be developed to compensate for the non-stationary nature of the brain signals.
A CRBM is a Generative Neural Network (GNN) which is a neural networks in which neurons’ state are updated according to the previous state of the network. These networks have been popularized by Hopfield in 1982 and extends to stochastic case with Boltzmann Machines. In a GNN, the state of interest in the equi librium state i.e. the convergence of the network after some updates. The diffusion network is a GNN in which the structure of neurons have been redesigned to store the dynamic of data. Such model have an attractive property: they can usually model a stochastic differential equation. For this reason, diffusion networks are particularly suitable for applications involving the context, and more particularly for processing temporal sequences such as learning and signal generation. Modelizaing brain signals as a stochastic signal is a particularly attractive idea which conduct to new applications of signal reconstruction and detection...
By realizing the developed algorithm in the Field-Programmable-Gate-Array (FPGA) embedded in the neural recording/stimulation system in NTHU, the ability of the algorithm to control DBS will be verified with animal experiments. Finally, the data recorded before and after closed-loop DBS will be analysed by statistical exploratory data analysis to decide the added value of thegraphical model: the evoked tools are analyse of variance, and high order statistics.


Objective
The PhD thesis aims to develop an evolutionary algorithm capable of reproducing the main stochastic dynamical properties of a complex system (e.g. brain circuitry) automatically. As the evolutionary dynamic model does not consist of physical equations like conventional models (e.g. Kalman model), the observability and controllability properties of the evolutionary dynamic model will be investigated.

Supervision:
The PhD student will be co-supervised by Pr. Vincent Vigneron, IBISC, Univ d’Evry in France and Pr. Hsin Chen, Elec. Engineering Dept., National Tsing Hua University (NTHU) in Taiwan. Pr. Vigneron is an expert in machine learning and signal processing, and Pr. Chen is an expert in probabilistic models and the hardware implementation of probabilistic models for biomedical applications. The PhD thesis aims to train a PhD student with interdisciplinary knowledge in machine, electronics, and basic neurophysiology. The student will focus on the development of machine learning in the 1st year. In the 2nd year, he will visit the NTHUconsortium to realize the developed algorithms as firmware for animal experiments. In the 3rd year, the data collected from animal experiments will then be analyzed to evaluate the effectiveness of the algorithms, as well as the underlying neuromodulation induced by the closed-loop DBS. The research findings of this PhD thesis are expected to be beneficial to not only improving the treatment on Parkinson’s disease but also facilitating the understanding on various brain circuitry.
Context
Pr. Vigneron has cooperated with Pr. Chen to develop machine learning algorithms for brain machine interfaces since 2008. They are currently cooperating with neuroscientists in the NTHU to develop an adaptive microsystem able to control deep-brain stimulation (DBS) in a closed-loop manner for treating the Parkinson’s disease. This research has being supported by the National Health Research Institute in Taiwan. The proposed PhD thesis is expected to contribute significantly this cooperative project by devising an adaptive system able to recognize pathological rhythms reliably regardless of the non-stationary nature of the brain, so as to enhance the reliability/efficiency of closed-loop DBS.


Expected results
1. An algorithm able to estimate the state of a complex dynamic system, as well as to infer the high-order structure of dynamics by exploring hidden-variable representations.
2. Enable the co-adaptive mechanisms of a hybrid system consisting of an artificial system and the brain.

Material and financial scientific conditions of the research project :
As part of the co-supervision, the university partner (NTHU) is committed to make available financement : to the student resource for campus housing, nourish during the 2nd year. Transport (2 ways during 1 year) and the fundings for 2 conferences in the 2nd year is also the burden of NTHU.


Objectifs de valorisation des travaux de recherche du doctorant : diffusion, publication et confi-
dentialité, droit à la propriété intellectuelle,. . .

This research work responds to the imperative of improving the management of PD on its health territory, while respecting the conditions of medico-economic efficiency. In fact, only 10% of European hospitals have the minimal infrastructure to receive Parkinson patients. Our goal in fine is to provide a response to needs in care improvement tools: linked to the homogeneity of treatments item access to care: the software will be shared to improve productivity: the calculation server is accessible remotely, speed of calculations ( The objective is to validate the results on multicentric patient databases and to integrate the model into
clinical application software with a comprehensible interface for the doctor.
The publications will be drawn up under a convention on conclusions protecting the authors of the con-
sortium and the intellectual property rights will be carefully examined for innovation. The protection will
be extended to the annotated database which will be used for learning and validation. A development
process to support the certification of the system will be put in place.
The supervisor will ensure that the research work will be published during the thesis without interfering
with the patenting process and the confidentiality of research.

Profil and searched skills: The person recruited must have an engineering degree or a Masters degree, solid knowledge in artificial intelligence, for example in deep learning (DL), in deep neural networks and in coding (Python, Cuda, C++). Experiences with graphics processor development (GPU) will be highly appreciated.
His English will be fluent. The work will be carried out at the IBISC Laboratory, a SIAM team on the premises of the UFR ST located on the Evry campus of the UPSaclay. IBISC develops multidisciplinary, theoretical and applied research in the field of information sciences and automatic learning, with a strong orientation towards health applications. The selected candidate will have the chance to work in an interdisciplinary team and with a consortium of data scientists and clinicians from the CHSF.

Contact: Vincent Vigneron, Hichem Maaref
(vincent.vigneron,hichem.maaref)@ibisc.univ-evry.fr
Phone: +33 6 635 687 60

For application, send CV+marksGPA+motivation letter

 


References
[1] Benabid, A. L. and Pollak, P. Mechanisms of Deep Brain Stimulation. Movement bibliographique Disorders 17, 73-74. 2002.
[2] Volkmann, J., 'Deep brain stimulation for the treatment of Parkinson's disease,' journal of clinical neurophysiology, vol. 21, no. 1, pp. 6-17, 2004.
[3] Beurrier, C., Congar, P., Bioulac, B., and Hammond, C., 'Subthalamic nucleus neurons switch from single-spike activity to burst-firing mode,' journal of neuroscience, vol. 19, no. 2, pp. 599-609, 1999.
[4] Chen, H. and Murray, A. F., 'A Continuous Restricted Boltzmann Machine with an Implementable Training Algorithm,' IEE Proceedings of Vision, Image and Signal Processing, vol. 150, no. 3, pp. 153-158, 2003.
[5] V. Vigneron and H. Chen. Sparse data analysis strategy for neural spike classification. Computational Intelligence and Neuroscience, 2014, Jul. 2014.
[6] Rémi Souriau, Vincent Vigneron, Jean Lerbet, Hsin Chen. Boltzmann Machines for signals decomposition. Application to Parkinson's Disease control. XXVIIème Colloque francophone de traitement du signal et des images (GRETSI 2019), Aug 2019, Lille, France.

 

Dans cette rubrique

(c) GdR 720 ISIS - CNRS - 2011-2020.