Home Research
Research
Cancel

Research

This is a description of my work including projects, conferences and papers

On sparsity and elastic consolidated penalties for parameter efficient transfer learning

Machine Learning in Clinical Neuroimaging - Charité – Universitätsmedizin Berlin - Department of Psychiatry and Neurosciences

This is a study of a new type of optimizer which reuse/eliminate efficiently pre-trained information transfer when fine-tuning deep neural networks, decreasing potential sources of overparameterization caused by fitting large models in small datasets. We also studied the effects of removing parameter redundancy and information preservation (Avoid catastrophic forgetting) in transfer learning settings.

This was my master’s thesis. Link: Soon (Consolidating paper and getting distribution permissions)

Harnessing the effects of continual regularization penalties to learn adversarial examples effectively

Research - Universität Potsdam

We found how continual learning regularization penalties such as Elastic weight Consolidation EWC could improve the generalization-robustness trade-off produced by the distribution mismatch between adversarial examples and clean data. This means when fine-tuning on adversarial examples continual learning regularization penalties could improve the generalization performance on the clean data lost by standard adversarial training strategies.

Link: Soon (Getting distribution permissions)

Adversarial

Enhanced Unsupervised-unpaired image to image translation for facial photo editing

Research internship - Universität Potsdam/Asaphus Vision GmbH

I developed a self-regularization facial recognition loss based on Neural Style Transfer and other improved GAN techniques to decrease CycleGAN identity deformation in facial attribute translation tasks.

Conferences

  • Bayesian Estimation of stock market Value at risk using VineCopula Models - M. Alba and W. Pineda - R/Finance 2018, University of Illinois, 2018 Link

  • Bayesian estimation of Value at Risk (VaR) for non-linear dependences in financial series in Colombia - M. Alba and L. Torres and D. Triana and W. Pineda - International Symposium of Statistics, IWAS, 2017, Medellin - Colombia

  • Vector Error Correction (VEC) model to evaluate the impact of macroeconomic variables on financial risk in Colombia (2010-2016) - M. Alba and L. Torres and D. Triana - International Symposium of Statistics, IWAS, 2017, Medellin - Colombia

  • Copula Theory - M. Alba - Research Poster Presentations: Universidad Santo Tomas, 2016, Bogota - Colombia

Papers/publications

  • Alba, M., L. Torres, D. Triana, and W. Pineda (2018). Bayesian estimation of Value at Risk (VaR) for non-linear dependences in financial series in Colombia - Comunicaciones en Estadística, vol: 11 (No.2),pp.171–189. (Currently only in spanish)