Francesco Innocenti

Postdoc @UniofOxford
| 🧠 PhD in ML & Theoretical Neuroscience @SussexNeuro | 🧑‍💻 @AmazonScience Intern

[Francesco Innocenti @InnocFrancesco]

🧠 Are brain-inspired algorithms inherently unscalable compared to backpropagation?

🎉 Pleased to share that our work on scaling predictive coding to 100+ layer networks has been accepted at NeurIPS 2025.

💻 Notebook: https://t.co/TDc3wsY9E5
📄 Paper: https://t.co/7UhbrbqNPq

[Anil Seth @anilkseth]

2/3 Francesco Innocenti @InnocFrancesco - a deeply suspicious (& brilliant) Ph.D. student @SussexUni @SussexAI - has now (w/ @drclbuckley) developed a new predictive coding method that scales to >100 layers w/ highly competitive benchmark performance

[Francesco Innocenti @InnocFrancesco]

🧠 Interested in local learning algorithms?

🚨 In work accepted at #ICML2026 🇰🇷, we show that predictive coding computes the same gradients as backprop for *much wider than deep* networks.

💬 Come chat at our poster (#802) today at 14:30! @mido7e

📃: https://t.co/D2SjBblkHh

[Francesco Innocenti @InnocFrancesco]

@lorenzo_noci Hey Lorenzo! I’m also at NeurIPS, it would be good to chat 🙂

[Francesco Innocenti @InnocFrancesco]

Interested in bio-plausible deep learning and in-context learning theory? 🧠 @mido7e I will be at NeurIPS this week to present 2 separate works:
- our "muPC" paper at the morning poster session on Thursday, &
- some ICL theory at the WCTD workshop on Sunday.

Come chat to us! 💬

[Francesco Innocenti @InnocFrancesco]

🧠 Are brain-inspired algorithms inherently unscalable compared to backpropagation?

🎉 Pleased to share that our work on scaling predictive coding to 100+ layer networks has been accepted at NeurIPS 2025.

💻 Notebook: https://t.co/TDc3wsY9E5
📄 Paper: https://t.co/7UhbrbqNPq

[Anil Seth @anilkseth]

2/3 Francesco Innocenti @InnocFrancesco - a deeply suspicious (& brilliant) Ph.D. student @SussexUni @SussexAI - has now (w/ @drclbuckley) developed a new predictive coding method that scales to >100 layers w/ highly competitive benchmark performance

[Francesco Innocenti @InnocFrancesco]

@auto_grad_ @mido7e Our extension is for *any* block as opposed to all blocks, such that if you take a block's input as given, you can apply the original result (nothing crazy).

I don't think this style of derivation can generalised to all blocks with some complex implicit update across layers.

[Francesco Innocenti @InnocFrancesco]

@neurosock Thank you! 🙏

[Francesco Innocenti @InnocFrancesco]

@jeffreyseely Thank you, I really appreciate it!

[Francesco Innocenti @InnocFrancesco]

@leothecurious Thanks!

[Francesco Innocenti @InnocFrancesco]

@anilkseth @UniofOxford Thank you Anil, for everything! 🙏🏻

[Francesco Innocenti @InnocFrancesco]

Lastly, @mido7e and I will be in San Diego to present this work, along with a workshop paper on the in-context learning capability of transformers (more on this soon). Feel free to get in touch! 🙂

[Francesco Innocenti @InnocFrancesco]

This paper concludes my PhD. None of this would have been possible without my amazing collaborator @mido7e, my supervisors @drclbuckley @anilkseth, and the support and funding of @SussexNeuro @RuthStaras. 🙏