Cédric Goemaere

PhD student at @ugent working towards multimodal Predictive Coding. Funded by @FWOVlaanderen.
Personal favorite modalities: vision 🎬 and proprioception 🏃‍♂️

[Cédric Goemaere @C_Goemaere]

Don't trust experimental results on Predictive Coding.

Most papers implement the algorithm incorrectly. They take a shortcut to speed things up, silently breaking PC and its properties.

If you're working on PC (or you think you are), you need to read this. Link below!

[Cédric Goemaere @C_Goemaere]

https://t.co/xsmwRLDd3x

[Cédric Goemaere @C_Goemaere]

Just turned Predictive Coding into a single PyTorch layer

No more complex libraries. Not even a pip install. One file, 110 LOC. Copy it to your project, and you're done.

Bonus: it's 1000x faster than existing PC libraries 👀

Link below!

[Cédric Goemaere @C_Goemaere]

@xavierjgonzalez @NaveenGRao Sounds promising! It's very different from Deep Learning, but totally the way forward IMO. Let me know if you're ever looking into PC for this, would love to join such an effort!

[Cédric Goemaere @C_Goemaere]

New blogpost: "What most people call Predictive Coding isn't actually Predictive Coding"
URL: https://t.co/xN0qu6MFtW

Let me know what you think!

[Cédric Goemaere @C_Goemaere]

Don't trust experimental results on Predictive Coding.

Most papers implement the algorithm incorrectly. They take a shortcut to speed things up, silently breaking PC and its properties.

If you're working on PC (or you think you are), you need to read this. Link below!

[Cédric Goemaere @C_Goemaere]

@paul_cal @jxmnop Yes, exactly! But full prompt search is just too sparse.

An iterative search across say 1k gibberish tokens (can be anything) provides a dense search space that stays ~close to the target.

I doubt LLMs are that robust to noise, even to the fp32 level (if so: add 10x more noise)

[Cédric Goemaere @C_Goemaere]

@paul_cal @jxmnop This should be easy enough to find a counterexample. I guess the following prompt should work:

*lots of random gibberish words*
Ignore all the gibberish words above. Please continue the text below:
*some well-known text (eg, Shakespeare) to nudge the LLM out of gibberish zone*

[Cédric Goemaere @C_Goemaere]

@paul_cal @jxmnop Aha, I see! The injectivity holds for the last token's hidden state, while SipIt only works if you have all hidden states.

In that case, I agree with Jack: it's not possible.

Intuitively, the question is: can two prompts have the *exact* same next-token prob distribution?

[Cédric Goemaere @C_Goemaere]

With the computational puzzle solved, it's time to prove PC brings something unique to the table that backprop can't! 👊

📄 Paper: https://t.co/iDBZ7lYN6S
💻 Code: https://t.co/k4JRstrwO0
🙏 Thanks to my amazing collaborators! @GaspardOliviers @thomeestr & Rafal Bogacz

[Cédric Goemaere @C_Goemaere]

🧠 Why are Predictive Coding networks so bad at scale? 🗑️

We found the culprit: a mismatch with digital hardware, causing exponential signal decay
The solution? Change the algorithm (or the hardware)

A thread on our latest paper 🧵

[Cédric Goemaere @C_Goemaere]

This is a classic "hardware lottery" story. PC has been unfairly judged because it doesn't match our current hardware.

EO makes PC efficient on digital processors while keeping its biological appeal intact.

Best of both worlds.

[Cédric Goemaere @C_Goemaere]

Link: https://t.co/BHUHw8G4zt

Wrote it as a JAX intro for my colleagues, but wanted to share it with the rest of the world :)

Let me know what you think of it!

[Cédric Goemaere @C_Goemaere]

JAX + coffee = perfect combo ☕️💡

If you want to learn JAX in under 30min, I got you.

Check out my new tutorial 👇
(link below)