Hi there 👋
I’m Emma, a machine learning engineer, data scientist and neuroscientist based in Barcelona ☀️
My work is about understanding how learning works, in biological brains and in artificial systems, and applying that to solve real-world problems with data.
I spent my PhD and postdocs modelling how the brain learns from limited experience, using reinforcement learning, signal processing, and dynamical systems. Since then I’ve been building and deploying machine learning models in production across a range of industries and data types: 3D point clouds, satellite imagery, geospatial data, biometric time series, and enterprise tabular data. I’m interested in roles that combine rigorous ML research or engineering with real-world impact, particularly in areas where scientific data, sensor signals, or complex structured data require thoughtful modelling approaches.
Currently, I’m an Applied ML Engineer at Visible, building algorithms to help people with energy-limiting chronic illnesses to pace and manage their condition.
→ See what I’m reading
I was awarded my PhD in Neural Dynamics in 2020, and completed two postdocs in computational neuroscience. My academic interests are in how the brain uses “downtime” (sleep, and rest when otherwise not actively engaged in behaviour) to generate spontaneous activity that supports interesting cognitive phenomena such as learning, forgetting, and generalising. These are also some of the big challenges in machine learning at the moment too: while techniques that are known to work in deep neural networks can serve as hypotheses for biological computational brain function, understanding better how nature has solved such problems can inspire more efficient machine learning algorithms.
In my PhD I used reinforcement learning models to understand how the brain “at rest” (sleeping, or quietly resting) contributes to offline learning and memory processes. I found that reward prediction errors are likely to influence which brain activity gets replayed during offline states.
📎 Paper
📎 Code
📎 Related code for identifying repeated patterns of brain activity (“replay”) using a naive Bayes classifier
Inspired by the surprising links between reinforcement learning in brains and reinforcement learning as a branch of artificial intelligence, I also wrote a review of how researchers in both fields might benefit from the advances in the other field.
📎 Paper
I have reviewed papers for Journal of Neuroscience, NeurIPS workshop on Biological and Artificial Reinforcement Learning, PLoS Computational Biology, and Cell Reports.
Take a look at some of the AI-for-good and data-for-good projects I’ve been involved in.
Early intervention in sustainability intelligence: a company launches a campaign to monitor the sustainability of its supply chain, inviting suppliers to get assessed by EcoVadis. EcoVadis proactively predicts how likely each supplier is to complete their assessment, targeting extra resources at those that need extra convincing. Partway through the campaign, there are signs that something is not right: the predictions indicate under-performance, threatening the client company’s sustainability programme. I revamped the binary classifier that makes these predictions, incorporating real-time signals to make the model more responsive, and used SHAP to help teams understand why campaigns were under-performing and what they could do about it - before the campaign had completed. This supported EcoVadis frontline teams and clients’ procurement teams to meet their supply chain sustainability targets.
Classifying risk construction elements: could you tell if a wall was built 5cm off from where it should be? At Naska.AI, I developed techniques to classify LIDAR point-cloud representations of architectural elements such as walls as being correctly positioned or not. I used classification filtering, hierarchical clustering based on learned representations, and vision-language foundation models to continuously improve on our core classification model. The result enabled clients to identify risk in their construction projects early, and reduce waste in one of the most polluting (but necessary!) industries.
📎 YouTube video about our software
World Data League 2023: my team (AI Wonder Girls) won first prize for our solution to promoting sustainable energy communities among residents of a Belgian city, using data to fight energy poverty and reduce carbon footprints. In earlier phases of the competition, we also provided solutions for understanding the biodiveristy of marine species and modelling traffic movements to reduce reliance on cars.
📎 Code
📎 3-minute presentation
DataKind UK & Evidence for Development: Evidence for Development is an NGO that maps data-driven “livelihood zones” to a country to improve the quality of interventions for development and disaster response. Mapping livelihood zones is very labour-intensive, so we used open-source data (remote sensing, transport and other geospatial data) to prototype an automated way to replace the first stage of this process, potentially speeding it up by a year or more per country.
📎 Blog post about the project
Omdena & GPSDD: GPSDD wanted to use open-source satellite imagery to predict the yield of several crops crucial to Senegal’s food security. Limited ground-truth data was a substantial limitation, so we trained a convolutional neural network on data from similar countries with better data and fine-tuned it on Senegal, combined with a random forest predictor to predict the yield at the end of the season. Users can use the forecasts to anticipate when poor crop yields might cause food insecurity, and take appropriate action.
📎 Project summary
This is the immodest part. Some of these awards are now quite old, but I appreciate them because they show off some of the skills that I am most proud of: my research skills, and my ability to communicate complex topics to different audiences.