Skip to main content Skip to navigation

Dr Julio Hurtado

 

I am an Innovation Research Associate (InRA) at CAMaCS and an International Collaborator with the Centro Nacional de Inteligencia Artificial in Chile (CENIA).

I did my PhD in Computer Science at the Pontificia Universidad Católica de Chile, specifically at the Laboratorio de Inteligencia Artificial (Artificial Intelligence Laboratory). Afterwards, I worked as a Research Associate at the Pervasive IA Lab (PAILAB) at the University of Pisa.

My main research focus is on Continual Learning, which aims to provide deep learning models with the ability to continually learn new knowledge without forgetting or interfering with previously learned weights. I am driven by learning representations more suitable for Continual Learning, encouraging generalisation and robustness in the model. My central hypothesis is that encouraging models to accumulate reusable and robust knowledge will help reduce forgetting by minimising the need to change previously learned weights. Related topics to the previous idea are Out-Of-Distribution generalisation, Open-World learning, robustness and learning disentangled representations.

Before and during my PhD, I worked as a Data Scientist on a few projects, applying machine learning techniques to real-world problems and transferring research findings into real applications using NLP and Computer Vision models.

I worked in a few start-ups. In the first one, we used NLP and Computer Vision models to categorise influencers on Instagram, improving connections between brands and users. The second involved a mobile agent that moved through supermarket aisles, analysing product prices and stock-outs. My job here was to develop a model capable of detecting different products to find when there was a stock-out problem or misplaced prices.

During my time as a PostDoc in Pisa, I also helped develop AvalancheLink opens in a new window, an End-to-End Continuous Learning Library based on PyTorch.

Current CAMaCS Projects:

  • Large Language Models for Rough Sleeping Policy Insights
    • Local authorities produce yearly reports on the successes and challenges of addressing rough sleeping. However, the significant volume of data makes it challenging to uncover insights and impactful directions that can benefit the reduction of rough sleepers.
    • In this project, we harness the power of large language models to index the data produced by local authorities in a vector database. This database is then queried with questions created by analysts, aiming to extract valuable information from the reports with answers generated by LLM. This pipeline is also known as Retrieval Augmentation Generation (RAG).
  • Revolutionizing Cataloging with Visual Transformers for Second-Hand Garments
    • The second-hand garment market is experiencing a surge in popularity. However, a significant challenge it faces is the wide variability in user-generated images, which significantly hampers searching for and comparing similar garments.
    • In this project, we seek to generate a catalogue of garment representations. Using pre-trained visual transformers, we explore different ways of generating representations, either with classical fine tuning or Parameter-Efficient Fine Tuning, like soft prompts or LoRA.

Recent publications

For a complete list: Google Scholar