Avatar

Mauro Camara Escudero

Research Associate in Statistical Machine Learning

University of Bristol

About Me

Passionate about leveraging data to drive positive change, particularly in tech-for-good sectors such as democracy, AI safety (particularly (mis-)alignment and training dynamics), energy, and environmental sciences. Pioneered the first algorithm for efficient Approximate Manifold Sampling and contributed to the development of Integrator Snippets, with a focus on Simulator-Based Inference (SBI). Some random things I have organised: Generative Models Reading Group, Neural Network Reading Group, Bristol Online Charity Danceathon, Postgraduates in AI (PAI) Link. Below, you can find my course on LLMs.

Other things I am passionate about include:

  • Basketball ๐Ÿ€
  • Dancing (mostly Modern Jive, West Coast Swing and Blues) ๐Ÿ’ƒ
  • Journalism, especially investigative (e.g. Bellingcat) ๐Ÿ•ต
  • Science of Cooking ๐Ÿงช ๐Ÿฅ˜
  • Exercising whether bodyweight, calisthenics or at the gym ๐Ÿ‹๏ธ

Interests

  • AI Safety, Trustworthy AI
  • Score-based sampling (HMC, MALA)
  • Generative Models (LLMs, DDGM)
  • Approximate Manifold Sampling
  • Scalable/Robust inference (SMC,SBI)

Education

  • PhD in Statistical Machine Learning, 2020-2023

    University of Bristol

  • BSc in Mathematics with Year in Employment, 2019

    University of Southampton

Projects

*

BBC BERTopic Modeling

Unsupervised Topic Modeling on BBC articles.

Dante GPT

LLM trained on Dante Alighieri’s Divina Commedia using Pytorch.

Exploring News Through Data

LDA & Sentiment Analysis on YouTube comments.

Maternal Mortality

A study on the lifetime risk of maternal mortality.

Spotify Wrapped Weekly

Real-time Data Visualization via the Spotify API.

Experience

 
 
 
 
 

Research Scientist Intern

Afiniti

Sep 2021 โ€“ Dec 2021 London
Designed a high-speed, high-accuracy low-rank treatment effect model in Julia, outperforming the company’s current approach by over an order of magnitude.
 
 
 
 
 

Data Scientist and Modeller

Uniper

Jun 2017 โ€“ Aug 2018 Nottingham
  • Led a high-value project with technical and client development components.
  • Developed gas turbine blades damage detection software in OpenCV.
  • Modelled and implemented back-end software in SQLServer and Cassandra.
  • Developed wind power forecasting models with Keras and Sklearn.
  • Designed gradient-free optimization methods to enhance wind farm layout.
  • Deployed bespoke Tableau dashboards to aid gas traders.
 
 
 
 
 

Summer Researcher

Dr. Richardson Giles, University of Southampton

May 2016 โ€“ Jul 2016 Southampton
Modelled and simulated particle hopping in solar panels with Python.

Ramblings

Tutorials, courses and various other rambling.

Large Language Models

AI Safety

Mathematical Machine Learning

Normalizing Flows Course

Approximate Bayesian Computation

Recent Posts

Sampling from a specific level set of a Gaussian

How to sample uniformly from a level set of a Gaussian

Measure Theory for ML, AI and Diffusion Models

Measure Theory for Machine Learning from Scratch

Holistic AI Workshop

On October 4th 2023, Roseline Polle (Responsible AI Auditor and Researcher at Holistic AI) and Sachin Beepath (AI Assurance Officer at โ€ฆ

Assessing a Variational Autoencoder on MNIST using Pytorch

Learn how to visualize the latent space and generate data using a VAE in Pytorch.

Minimalist Variational Autoencoder in Pytorch with CUDA GPU

Variational Autoencoders in Pytorch with CUDA GPU

Talks

Selected talks during my PhD.

Approximate Manifold Sampling

Sampling from distributions concentrated around a manifold.

The THUG sampler

The first bespoke sampler for filamentary distributions.

Contact

  • Woodland Road, Bristol, BS8 1TH
  • Turn right at the entrance to office GA.14