Alexandra Gastone Guilabert

About

A neuroscientist and researcher by training, I'm passionate about working at the nexus of different disciplines. Currently, I'm continuing the research from my MSc studies with the Institute of Neuroinformatics (Zürich), and developing several other data science projects in sectors that include neuroscience, disaster relief efforts, sentiment extraction from social media, and credit risk analysis. For a look at some of my academic work, check out my Bachelor's thesis proposal from New York University along with the report we published in Nature Neuroscience for that project, and an abstract of my Master's thesis in ETH Zürich and University of Zürich for the Swiss Society of Neuroscience 2019 Annual Meeting.

Simulation of Cortical Neurons

MSc Thesis with ETH Zürich and University of Zürich: Synaptic background activity and short-term plasticity optimize synaptic information transfer between Layer 2/3 cortical neurons.

Credit Risk Predictor and Explainer

Predicts risk class of credit card applicants. Additionally generates a dashboard to visualize feature importances, classification results, and contribution to individual predictions. For more details, check out the Github repo.

Flight Finder

Webapp that searches, computes, and displays the lowest cost common flight destinations given two origins. Built in collaboration with Lara Riparip, colleague from ETH Zürich. Check out the backend and Lara's frontend, or visit the website here to see the progress so far!

Disaster Response Classification with BERT

Finetunes a pretrained BERT Transformer model on labeled disaster messages from social, direct, and news sources for multilabel text classification (more details on the training dataset by running the D3.js co-occurrence matrix visualization I built). Using Flask, deployed the model through a web app for real-time inference of user queries. Check out the project repo here for the model, the visualization of the training data, and instructions to run the web app.

Sentiment Analysis on Airbnb reviews for a Collaborative Recommender System

Generates personalised recommendations for Airbnb users based on sentiment polarity of their and other users' comments. For more details, check out the Github repo, my article on Medium for an accessible guide to a simple implementation, this dashboard I created with Streamlit (deployed on Heroku) on estimated sentiment polarity data, or this Tableau dashboard for polarity and review scores by location.

Tweet Sentiment Extraction: a Question Answering Model with ALBERT

Uses Hugging Face's implemention of the ALBERT Transformer model on a Question Answering task, fine tuned on a custom SQuAD-like dataset of tweets, to extract the words that best support a positive, negative or neutral sentiment label. Currently evaluates to a score of 0.7114 (top submission at time of writing is 0.724). Check out the project.

Anomaly Detection of Lorenz Attractor model data

Trained a feed forward neural net (Tensorflow + Keras) on extracted frequency spectrum data of a Lorenz attractor model to detect anomalies (supervised). To address the lack of labeled anomalies in these types of problems, additionally trained a LSTM autoencoder on healthy data and set an error threshold for the loss above which a data point will be classified as an anomaly (see Github repo).

Simulations of Auditory, Visual, and Vestibular Sensory Systems

Computer simulations of a cochlear implant, a retinal implant, and a vestibular implant (see Github repo) implemented in Python with a custom-built GUI for each simulation. Completed in collaboration with colleagues from ETH Zürich.