The work spans machine learning, neuroscience, and scientific modeling, with a consistent technical
agenda: build systems that generalize reliably and quantify what they do not know.
Working Statement
I am an independent researcher with training in computational neuroscience, machine
learning theory, and scientific modeling. Much of the recent work is about uncertainty,
regularization, and the geometry of learning; the older and still active threads run
through plasticity, motor control, color, and representation.
uncertainty-aware ML
learning theory
computational neuroscience
adaptive control
Research Areas
A compact map of the work.
These threads fit together around a shared set of questions: how learning proceeds under limited
evidence, how uncertainty can track reality, and how structured systems adapt without collapsing
into brittle heuristics. The same lens also extends to scientific discovery itself when agents
become collaborators in reasoning, search, and experiment design.
Uncertainty and calibration
Methods for making models express what is genuinely supported by data. This includes
calibration-aware objectives, Bayesian neural networks, and uncertainty for neural PDE
surrogates.
Learning theory and geometry
A theory of learning time, finite-data thresholds, signal-to-noise, curvature, and high-dimensional
effects in optimization and generalization.
Motor control and world models
Cerebellar-style controllers, embodied reinforcement learning, and reference-trajectory world models
for fast adaptation under changing dynamics.
Plasticity, perception, and representation
Local learning rules, cortical representation, correlation-invariant synaptic plasticity, and a more
speculative line of work on color and associative structure.
AI science and scientific discovery
Research on how agent systems can support inquiry itself: literature synthesis, hypothesis generation,
uncertainty-aware reasoning, and the structure of human-AI scientific collaboration.
Selected Highlights
A few anchor points.
Reliable AI
Cross-regularization, Twin-Boot, and precise Bayesian neural networks all address the same practical
issue: models should know when they are guessing.
Brains and control
Recent work uses motor adaptation as a meeting point for reinforcement learning, control theory, and
cerebellar computation.
Technical depth
Background spans theoretical neuroscience, deep learning, scientific computing, competitive programming,
and software engineering.
Other Work
Installations, colour, and consciousness.
Another part of the work centers on multi-agent public installations, colour perception, qualia,
and the structure of experience.
Chatsubo: AI bar
A live multi-agent social simulation built as an AI bar for the 2024
Metamersion: Healing Algorithms
exhibition in Lisbon. Autonomous bartenders, ghost patrons, memory, rumor, and human visitors all
share the same evolving social environment.
Colour theory and consciousness
A research line on colour qualia as learned associative structure, including
The Blue is Sky,
work on empiricist theories of consciousness, and experiments on qualia drift under altered spectral environments.
Research Program
Selected papers and active threads.
Selected publications across learning theory, reliable AI, motor control, and computational
neuroscience. Links point to conference, journal, or arXiv records.
The uncertainty work is not cosmetic calibration. It treats uncertainty, regularization, model size,
and robustness as parts of the same generalization problem.
Learning as a geometric process
A central question is why some structures are learned quickly while others remain slow or unreachable,
and how that depends on dimension, curvature, and finite data.
Controllers inside learned systems
Motor-control work separates long-horizon policy learning from fast corrective control, both in robots
and as a theoretical picture of cerebellar function.
Brains as learning algorithms
The neuroscience line asks which local plasticity rules can plausibly learn structure from raw sensory
inputs, and what that says about cortex and representation.
Science as a learning system
Another active thread asks how discovery changes when agents can read, compare, critique, and help
structure scientific reasoning without flattening the human part of the process.
Consciousness and Public Work
Colour, qualia, and live systems in public.
Another active branch of the work connects philosophy of mind, colour perception, and public-facing
agent systems. It sits closer to consciousness research and experimental aesthetics than to standard ML.
Academic history across research, service, outreach, and technical work.
Positions
2025-present
Independent researcher, Lisbon
Self-directed independent lab (NightCity Labs) researching on trustworthy AI, uncertainty-aware machine learning, AI for science,
computational neuroscience, and adaptive systems.
2022-2025
Research Scientist, Champalimaud Research
Natural Intelligence Lab, Champalimaud Centre for the Unknown, Lisbon.
2020-2021
Machine Learning Expert, Cambridge Spark
Course design, large-scale machine learning training material, and the G-Research Kaggle competition.
2020-2021
Visiting Scientist, EPFL
Laboratory of Computational Neuroscience, Lausanne.
2016-2019
Postdoctoral Researcher, Gatsby Computational Neuroscience Unit, UCL
Postdoctoral work in theoretical neuroscience and machine learning.
2006 / 2003
Engineering internships at Google and IAE/CTA
Software systems, optimization, and computational engineering.
Education
PhD in Computational Neuroscience, EPFL, 2010-2016.
MSc in Neuroscience, University of Sao Paulo, 2008-2009.
BSc in Computer Engineering, ITA, 2002-2007.
Awards
Silver Medal, International Mathematical Olympiad (IMO), 2001.
World Finalist, International Collegiate Programming Contest (ICPC), 2005 and 2006.
Gold Medal, Brazilian Mathematical Olympiad, 2000, 2001, 2003, and 2004.
2nd prize, International Mathematics Competition, 2004 and 2005.
First prize, Brazilian Physics Olympiad, 2001.
Selected talks
Generalization, regularization and Bayesian neural networks. Gatsby UCL Transfer Talk, 2019.
Adaptive regularization by noise in deep neural networks. Albert Einstein Hospital Big Data Initiative, 2019.
Five ways to regularize deep neural networks. LCN EPFL Seminars, 2018.
Theory of synaptic plasticity for receptive field development. Oxford Neurotheory Seminars, 2015.
Founder and organiser, Theory Mini-Symposium series, 2022-2024.
Referee for Cosyne, NeurIPS, ICML, and ICLR.
Guest editor, PLOS Computational Biology.
Outreach
Science on the Walls | Ciencia nas Paredes, Lisbon, 2024. Artificial Intelligence and Medicine,
Itau Cultural, Sao Paulo, 2018. How the brain represents the world, Arte della tavola, Lausanne, 2013.
Languages and tools
Portuguese, English, Spanish, French, and German. Technical stack includes Python, C++, Julia, Matlab,
PyTorch, and agentic systems tooling.