Research Interests#

My focus in on the application of Machine Learning methods in the areas of Smart Cities, Neuroscience, and High-Performance computing.

Smart Cities

  • Improving the Quality of Public Bus Systems

Public bus systems are an essential part of urban mobility, due to their low cost and high capillarity. But providing acceptable for users is difficult due to limited system capacity, interactions with the city traffic, and disruptions caused by city events. We analyze GPS from public buses from São Paulo city to better understand the factors that impact the reliability of bus systems, to determine disruptions in the bus systems in real-time, and to predict the future states of the bus systems. We are also working with the problem of the last mile, since it normally incur high-cost, and a poor quality of service. We are evaluating incentives for people to use multiple types of models, including e-hailing services.

  • Using Public Health Data to Improve Public Policies

EEG Decoding

  • Deep Learning for EEG Decoding in BCI and Sleep Staging

EEG signals are generated from the collective activity of billions of neurons and allow us to monitor brain function using non-invasive electrodes. The captured signals are noisy and contain overlapped signals from many regions, making them difficult to interpret. We use techniques such as transfer learning and self-supervised learning in deep neural networks to improve the decoding in EEG signals, in applications such as Sleep Staging and Brain Computer Interfaces (BCI).

High-Performance Computing

  • Learning better HPC Scheduling Policies

Existing supercomputers must process many thousands of jobs (programs) submitted by users. Defining how to prioritize the execution is the main task of job resource managers. We use machine learning techniques to improve this scheduling process by: (1) generating custom policies for specific architectures and workloads, and (2) selecting the best policy to apply to each machine given its current state.

Previous Research Topics

  • Accelerating Biology and Neuroscience applications using GPUs

Many biological and neuroscience applications require very large amounts of computational power and can benefit from the massively parallel processing units available on GPUs. We developed algorithms in CUDA for simulations Hodgkin-Huxley networks, determination of gene regulatory networks and enumerating hitting set solutions.

  • Functional Connectivity in Neural Networks

The brain contains thousands of regions that must communicate in an organized way. Directed Functional Connectivity can be estimated using causality methods, such as Granger Causality and Partial Directed Coherence. We use computational models of mice connectomes to evaluate the applicability of these measurements in neural signals.