MSc. in Computer Science, University of Concepción (UdeC)
Publications
2022
M. Pérez-Carrasco, B. Karelovic, R. Molina, R. Saavedra-Passache, Pierluigi Cerulo, and G. Cabrera-Vives Precision silviculture: use of UAVs and comparison of deep learning models for the identification and segmentation of
tree crowns in pine crops.
International Journal of Digital Earth (IJDE), 2022
[Paper][BibTex]
2021
M. Pérez-Carrasco, P. Protopapas, , and G. Cabrera-Vives. Con$^{2}$DA: Simplifying Semi-supervised Domain
Adaptation by Learning Consistent and Contrastive Feature Representations.
Published in the Distshift Workshop of the 35th Conference on Neural Information Processing Systems (NeurIPS) Workshop on Distribution Shifts, 2022.
[Paper][BibTex]
P. Sanchez-Saez, et al. Searching for Changing-state AGNs in Massive Data Sets. I. Applying Deep Learning and Anomaly-detection Techniques to Find AGNs with Anomalous Variability Behaviors.
Published in Astronomical Journal.
[Paper][BibTex]
F. Förster, et al. The Automatic Learning for the Rapid Classification of Events (ALeRCE) AlertBroker.
In press. The Astronomical Journal (AJ), 2021.
[Paper] [Repository] [BibTex]
2020
M. Pérez-Carrasco, G. Cabrera-Vives, P. Protopapas, N. Astorga, and M. Belhaj. Matching Embeddings for Domain Adaptation.
ArXiv 1909.11651, 2020.
[Paper] [BibTex]
2019
M. Pérez-Carrasco, G. Cabrera-Vives, M. Martinez-Marín , P. Cerulo, R. Demarco,P. Protopapas, J. Godoy, and M. Huertas-Company. Multiband galaxy morphologiesfor CLASH: a convolutional neural network transferred from CANDELS,
Publications of the Astronomical Society of the Pacific, 2019.
[Paper] [Repository] [BibTex]
Conferences and Workshops
Astroinformatics 2019. Expositor. Best student paper award. Caltech, Pasadena, CA, USA. Video.
ComputeFest 2019. Trainer. Harvard University, Boston, MA, USA. Repository.
Big Data Astronomy Workshop 2018. Expositor. University of Concepción, Concepción,Chile.
Teaching
Department of Industrial Engineering, @UdeC
Lecturer
This is an elective course aimed at graduate and undergraduate students of engineering and teaches some fundamentals on deep learning.
I co-taught with Professor Rodrigo de la Fuente.
At the beginning of the course we study neural network architectures such as MLP, CNN, and RNN; Then we move to more advanced topics such as reinforcement learning, representation learning, and deep generative models. The main goal of the course is to give a theoretical perspective, and a practical set of tools to prepare the students for facing large-scale real-world problems both in industry and academia.
Department of Industrial Engineering, @UdeC
Lecturer
This is an elective course aimed at undergraduate students of engineering and teaches the basics on machine learning.
I co-taught with Professor Rodrigo de la Fuente.
The course starts with a brief review of programming, linear algebra and statistics; In the second part we move to supervised learning (Regression, Logistic regression, decision trees, Random Forest, SVM) and unsupervised learning (k-means, PCA). The main goal of this course is to introduce machine learning algorithms for facing problems both in industry and academia.
Institute for Aplied Computational Sciences, @Harvard U.
Teaching Fellow
This is a mandatory course of the Master in Data Science. In this course I was in charge of office hours, grading and guiding projects.
The lessons introduces advanced methods for data wrangling, data visualization, deep neural networks, and statistical modeling. Topics include big data and database management, multiple deep learning subjects such as CNNs, RNNs, autoencoders, and generative models as well as basic Bayesian methods, nonlinear statistical models and unsupervised learning.
Department of Industrial Engineering, @UdeC
Teaching Assistant
This was an elective course aimed at undergraduate students of engineering. In this course I was in charge of recitations/laboratories.
This course mixed classical machine learning with deep learning algorithms. I showed how to use supervised and unsupervised learning for dealing with large-scale problems, with an special focus in industry issues.
Department of Industrial Engineering, @UdeC
Teaching Assistant
This is a mandatory course for 2nd year students of Industrial Engineering. In this course I was in change of recitations.
Systems Modelling teaches how to mathematically model and solve some basic operation research problems using linear programming. The objective of this course is to give an introduction for more advanced operation research/optimization courses.
Masters Thesis
I got my Masters degree in Computer Science at University of Concepción (UdeC). My advisors were Guillermo Cabrera (UdeC) and Pavlos Protopapas (Harvard U.). They both brought me strong support during the process.
How can deep learning models learn from few examples?
Deep learning models have demonstrated to be capable to learn underlying patterns in the data, becoming the sate of the art for a lot of machine learning problems. However, they face overfitting problems when the amount of data is scarce. Furthermore, learning theory is well defined only when train and test data comes from the same distribution and feature space.
Sometimes, there are vast annotated sources, which are similar to our few labeled target data, but have different distributions. In this cases it is possible to use domain adaptation as a tool for aligning both distributions into a common shared space, allowing us to avoid the overfitting, therefore improving generalization and performance on the target domain.
In my master thesis, I propose a method that uses deep generative models, specifically Deep Variational Embeddings (VaDE) and Generative Adversarial Networks (GANs), modeling a latent space composed by a Gaussian mixture model, where each label class corresponds to a Gaussian mixture component, independently of the domain membership.
Using this method, we achieved impressive results in low-dimensional benchmarks and a real-life scenario classifying images of galaxies by their morphologies.
This work was presented in the Astroinformatics 2019 conference at Caltech, being honored with the Best Student Paper award.
[Presentation] [Preprint]
Acknowledgements
My thesis was done 6 months at UdeC and 6 months at Harvard U., and It has been by far the most challenging and exciting research experience I ever had. That makes it an indelible memory.
I am very grateful to my advisors and people from UdeC, Harvard, MIT, and Brandeis who were part of this process. Thanks you!