Titulo Estágio
Model identification of epidemics via Deep Learning
Áreas de especialidade
Sistemas de Informação
Local do Estágio
CISUC-LARN
Enquadramento
The analysis of epidemic models provides the means to forecast and control the evolution of a pandemics. Several attributes play a major role in the faith of a pandemics: vaccination policy, mutation rate, cordon sanitaire policies and so on. Unfortunately, the qualitative behavior of these dynamical systems critically depends on the underlying dynamical laws of the models. This implies that accurate models are necessary to predict more precisely the degree of infection of a population over time and to devise optimal mitigation strategies as vaccination and cordon sanitaire policies.
In this internship, we resort to Deep learning methods, as convolutional and recurrent neural networks, in order to learn the dynamical laws (or the building blocks therein) governing the evolution of an epidemics from the observed risk map samples. More concretely, given a collection of samples reflecting the fraction of infected individuals over time across distinct regions of the globe, we would like to develop Deep learning based algorithms to consistently learn the dynamical law underlying the evolution of the pandemics.
In this internship, the student will be exposed to three hot topics of current research: epidemics, model identification/topology inference and Deep learning methods.
Objetivo
In this internship, the student should study, propose, implement, and test methods for model identification of epidemic dynamical systems via deep learning methods. In particular, the student should explore certain regression formulations of the model identification problem, i.e., formulating it as the minimization of a loss function whose minima is achieved at the ground-truth dynamical model parameters (consistency). This formulation will be tailored to epidemic models. Then, we will embed Deep Neural Networks into the regression formulation and resort to Deep tools such as stochastic gradient descent, back propagation, convolutional and recurrent neural networks to tackle the optimization problem and finally learn the underlying epidemic dynamical law (our ultimate goal).
To achieve this goal, the following objectives will be pursued:
- Study the state of the art regarding model identification of nonlinear dynamical systems and Deep learning;
- Propose potential model identification strategies that could apply to epidemic models (regression formulation);
- Implement these algorithms via Deep learning tools;
- Deploy test setup and study the performance/consistency of the algorithms developed.
Plano de Trabalhos - Semestre 1
- Study Deep learning methods such as stochastic gradient descent, back propagation, convolutional and recurrent neural networks;
- Explore Deep learning tools to address the model identification problem applied to epidemics (i.e., to address the regression formulation);
- Start implementing the proposed approach/algorithm with synthetic data;
- Write an intermediate report.
Plano de Trabalhos - Semestre 2
Validation of the methods developed with a real dataset:
- Implement the proposed solution and study its performance/consistency via numerical simulations or formal analysis;
- Test and evaluate its performance with a real dataset;
- Write the final report.
Condições
This work should take place in the context of a research project funded by FCT in a CISUC lab. There is the possibility of a 6-month scholarship of 835,98 euros per month.
Orientador
Augusto Santos / Catarina Silva / Bernardete Ribeiro
bribeiro@dei.uc.pt 📩