Titulo Estágio
GA-Assisted Deep AutoEncoders
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
CISUC
Enquadramento
Motivated by the increase in the computational power, deep learning has led to a widespread enthusiasm in the Pattern Recognition and Machine Learning areas. Deep architectures are inspired in the neurosciences and aim at decomposing a hard problem into simpler tasks, by stacking several layers of neurons, each one responsible for learning a different set of features. However, the problem of finding the appropriate topology for an Artificial Neural Network (ANN) often follows a trial-and-error approach, which is a difficult and time consuming task. In order to overcome this challenge, researchers have focused their attention on the development of algorithms to automate the discovery of adequate topologies (and/or weights) of ANNs relying on the use of Evolutionary Computation.
The current proposal is aligned with the research agenda of CISUC’s on Evolutionary Machine Learning, involving the Computational Design and Visualization Lab. and Laboratory of Artificial Neural Networks, who c
Objetivo
The main goal of the Thesis is the study, understanding and development of an evolutionary framework for the evolution of the topology of Deep AutoEncoders. In brief words, AutoEncoders are unsupervised learning models that aim at rebuilding the original data, i.e., whereas typical feed-forward neural networks try to predict the correct classes (y) from the input data (x), autoencoders try to predict x from x. Therefore, one of the main advantages of AutoEncoders is their capability to learn compressed representations of the original data, in a process that resembles similarities with common feature selection techniques, such as the Principal Component Analysis (PCA) algorithm. Then, the developed framework should be applied to well established benchmarks and the results compared with the ones achieved by existing approaches.
Plano de Trabalhos - Semestre 1
Task 1 - Revision of the bibliography and survey of the state of the art;
Task 2 - Determination of the techniques to use and develop;
Task 3 - Proof of concept and prototype development;
Task 4 - Writing of the intermediate report and refinement of the work plan for the second semester.
Plano de Trabalhos - Semestre 2
Task 5 - Development of the framework for the automatic evolution of AutoEncoders;
Task 6 - Experimentation and Analysis;
Task 7 - Writing of the the Thesis;
Task 8 - Scientific paper.
Condições
Strong skills in programming.
Will to communicate in English with other researchers is also important.
Other interesting skills include Complex Systems and Artificial Intelligence.
Observações
The candidate will have access to state of the art servers in order to run the necessary experiments. Additionally, he/she may also be granted a full research scholarship on the second semester, depending on the conducted work and funding availability.
Orientador
Penousal Machado & Bernardete Ribeiro
machado@dei.uc.pt 📩