Francesco Innocenti

Francesco Innocenti

PhD student in Theoretical Neuroscience

About me

Hi! I’m Francesco.

I’m currently a PhD student in Theoretical Neuroscience and Machine Learning at the University of Sussex, supervised by Christopher Buckley and Anil Seth . My PhD is about scaling a brain-inspired local learning algorithm called “predictive coding”, as a potentially more efficient alternative to standard backpropagation. I use both theory and experiment, with a focus on optimisation theory. In trying to better understand this algorithm, I have developed expertise in (and a fascination with) the optimisation dynamics and scaling laws of standard deep neural networks.

I recently interned as an Applied Scientist at Amazon, helping to improve their model forecasts to deliver packages throughout Europe more efficiently. During my undergrad, I worked as a research assistant in Ashok Jansari’s lab at Goldsmiths, University of London, helping to develop a face memory test aimed at people with both poor and exceptional face recognition abilities. I also interned in the lab of Devin Terhune, using machine learning tools to differentiate the subjective experiences associated with different psychedelic drugs.

Outside of work, I enjoy running, swimming and playing football, and am always up for a physical challenge!

News

💻‍ Introducing JPC

December 2024

⭐️ Our lab just released JPC , a JAX library for training neural networks with Predictive Coding. See this pre-print for more details.

🇨🇦 Paper accepted at NeurIPS 2024

September 2024

Our paper Only Strict Saddles in the Energy Landscape of Predictive Coding Networks? has been accepted at NeurIPS 2024! See this blog post for the key takeaways.

🇪🇸 Applied Scientist Intern at Amazon

April 2024

I recently completed an internship as Applied Scientist at Amazon, in Barcelona. I wrote about my experience here if you’re interested.

🏆 Best Workshop Paper Award at ICML 2023

June 2023

I will be at ICML 2023 in Hawaii to present my work on Understanding Predictive Coding as a Second-Order Trust-Region Method, which won a Best Paper Award at the Workshop on Localized Learning. You can rewatch my talk here or read my blog post for the key ideas.