About Me

Fabio Ferreira is a Postdoctoral Researcher in Frank Hutter’s Machine Learning Lab at the University of Freiburg. His research interest lies in exploring AI-generating algorithms and world simulations to uncover the fundamental building blocks of intelligence. He sees meta-learning, compression, bottom-up methods, and open-ended learning as key ingredients for artificial systems to evolve and innovate much like nature does.

Fabio received his Ph.D. in Computer Science from the same lab, where his dissertation focused on meta-learning for synthetic data generation and automated model and hyperparameter selection. During his PhD, he interned at AWS AI, developing meta-learned models for cost-aware performance prediction of ML algorithms. Prior to that, he worked in robotics as a student researcher in the Stanford AI Lab, focusing on learning physics models for robotic manipulation, and completed his CS studies at KIT, Germany, where he already explored action-conditioned video prediction models in 2017 at the H2T Lab.

Publications

For an up-to-date publication record, please check out my Google Scholar profile.

Coding and Open-Sourcing

• Paper implementation “Information Theoretic MPC for Model-Based Reinforcement Learning” (Williams et al., 2017)

• cde, Python package (with over 3k downloads per month) for conditional density estimation with neural network-based, non/semi-parametric and normalizing flow estimators, various data generating processes (AR(1), ARMA jump diffusion, GMM etc.) and evaluation metrics (KL and JS divergence, hellinger distance, etc.)

• videofeatures, Python package (pip install videofeatures) for video/image feature extraction (ResNet, VGG, SIFT, SURF) that, based on the selected feature type, trains a fisher vector GMM and computes (improved) fisher vectors

• video2tfrecord, Python package (pip install video2tfrecord) that allows easy conversion of RGB video data to the TensorFlow tfrecords format

• Planar Manipulatior Dataset, goal-oriented 2D toy dataset for deep neural network testing

• Flying Shapes Dataset, goal-oriented 2D toy dataset for deep neural network testing

• Paper implementation “Movement Primitives via Optimization” (Dragan et al., 2015)