San Francisco Bay Area
Wearable health-tech enthusiast and ML/AI practitioner in the wearable space. Previously a Senior Machine Learning Engineer at All.Health (a wearable health-tech company providing continuous monitoring of patients vitals -- formerly known as Jawbone). I developed ML prototypes, and deployed algorithms to production to improve patients' access to personalised health care. Formerly a research scientist at University College London, I developed and applied machine learning models to better understand the cognitive processes involved in Mental Disorders. My research aimed to uncover unique individual computational markers (computational cognitive fingerprints) explaining patients biases, decisions and predicting future symptoms. These model predictions could then be used provide early forms of treatment to prevent the onset of chronic mental health disorders. Skills and topics used in my research are: Computational modelling, Bayesian modelling & statistics, Reinforcement learning, Deep Learning, Neuroscience, Artificial Intelligence, Sampling, Human Behaviour, Computational neuroscience, Computational psychiatry.
Developing ML prototypes, and deploying algorithms to production at a wearable health-tech company. Modelling: Deep Learning (PyTorch, TensorFlow), Bayesian Statistics (Stan, PyMC), Predictive modelling & traditional ML (Scikit learn, XGBoost), Experiment tracking (MLFlow), Performance and throughput optimisation (Quantization) Deployments: Containerisation (Docker), Pipelines (Airflow), Model serving & versioning (ONNX • TensorFlow • MLFlow), Monitoring (Sentry • Prometheus • Grafana), Microservices development (Flask • Django • FastAPI • Kafka • Rabbitmq), Databases (PostgreSQL • TimesaleDB • MongoDB), Web services (GCP • Azure • AWS)
Awarded a 2-year independent NIHR-UCLH BRC Fellowship worth £252,264.52 (~350K$) to use machine learning on to better understand, diagnose and predict future mental health disorders. Developed Reinforcement-Learning models, Bayesian models, Drift Diffusion models and Hierarchical multilevel probabilistic models of decision-making to: 1 - Discover whether patients with Depression and healthy controls may be using different strategies or heuristics (models) to solve behavioural tasks (decision making) 2 - Uncover computational markers (latent parameters of the models -- computational fingerprints) for each individual participant (e.g. how sensitive to reward or punishment each patient is, or how fast patients update their internal model when presented with surprising novel information – learning rate –, etc.) 3 - Identify individuals at-risk of developing chronic mental health conditions based on extracted computational-markers (which may be an indication of suboptimal cognitive processing), in order to provide early forms of treatment to prevent the onset of disease
Worked with: Jonathan Roiser (Institute of Cognitive Neuroscience, UCL) & Peter Dayan (Gatsby Computational Neuroscience Unit, UCL) to develop computational analyses aimed at better understanding sub-optimal cognitive processes in mental health (particularly in patients suffering from Major Depression and Anhedonia). Computational analyses (ML/AI models): Developed hierarchical multi-level models of cognition and decision making using probabilistic programming (HMC Sampling) in Stan (e.g. reinforcement-learning models, Bayesian models, Drift Diffusion models). Performed data analysis, classification and solved optimisation problems on medical research data. Experimental design: Designed behavioural tasks to assay various aspects of decision making in humans: - reward & punishment processing, - biases - motivation, - mood, - resilience to failure / helplessness
Created teaching material in the form of python tutorials for 1700+ interactive students and 5000+ observer students taking part in the Neuromatch Academy Summer School 2020 in Computational Neuroscience. Tutorials covered: Bayesian statistics, Bayesian inference, cost functions, decision theory, and model inversion. A sample of the teaching material created for the course is available at: https://github.com/NeuromatchAcademy/course-content/tree/NMA2020/tutorials/W2D1_BayesianStatistics
• Designed 16h+ of Python teaching material for 40+PhD students Topics included: Optimisation, Model fitting, Model comparison, Signal Detection Theory, Drift Diffusion Modelling, Bayesian models of cognition, Computational Psychiatry • Co-tutor for the reinforcement learning module Samples of tutorials is available at: https://github.com/vincentvalton/CCNSS_2018_tutorials