MKoudahl

@exilefaker Solve for argmax p(delicious | prep)

A Bayesian approach to breakfast - Sample dish from p(dish) - Sample preparation method from p(prep | dish) - Sample 2 utensils from p(utensil | dish) - Make the best of it

@lancelotdacosta @VERSESAI AWESOME!!! Well done my friend 🤘🤘🤘

BIG UPDATE: ✅ I finished my PhD entitled ‘Probabilistic principles for biophysics and neuroscience’ ➡️ I now research #structure #learning @VERSESAI 🙌🏼 I currently visit Yoshua #Bengio to learn about #GFlowNets 🚀 An incredible journey lies ahead! 🙏🏻👌🏻

Link for those interested :) tinyurl.com/mve6jth8

Happy new years! To start 2024 off right, I'm becoming a doctor! The defense ceremony is public and everyone is invited! If anyone wants to join in Eindhoven, you are so welcome! If you'd like to watch me get roasted but can't quite make it, there's gonna be a public livestream

Also FUCK YEAH!!!

🚨 Preprint 🚨 “Active inference and intentional behaviour,” from @VERSESAI (Friston, @TommSalvatori, @a_tschantz, @exilefaker, Verbelen, @KoudahlM, @aswinpaule, @drclbuckley, me) and collaborators (@TakuyaIsomura, Parr, @adeelrazi, @ANeuroExplorer) arxiv.org/abs/2312.07547 1/8

@BloodBowl_Game Necromantic Horrors

"🏈 Tomorrow, we unveil the newest faction joining Blood Bowl 3 in Season 3! 🌟 Can you guess which one it is? A winner will be drawn from the correct answers and will win 1000 Warpstone – a perfect kickoff to the season! 🚀

🚨 Preprint 🚨 Stoked to share “Supervised structure learning,” from @VERSESAI authors (Friston, @lancelotdacosta, @a_tschantz, @exilefaker, @TommSalvatori, Koudahl, @conorheins, Verbelen, @drclbuckley) and collaborators (Neacsu, Sajid, Markovic, Parr) arxiv.org/abs/2311.10300 1/5

Our new preprint led by the great Karl Friston, in which we explore how the structure of generative models can be learnt from data arxiv.org/abs/2311.10300

@BloodBowl_Game Wood elves

🎁 The next Bloodbowl Faction will be revealed in just a few days! Take a guess in the comments for a chance to win 1000 Warpstone! 📅 3 Winners will be selected on the 21st.

Honored to announce that I'll be writing a book on the FEP and active inference for the @mitpress, tentatively entitled: (Artificial) Intelligence from First Principles: Using the Free Energy Principle and Active Inference for the Study and Design of Ecosystems of Intelligence

@fchollet @karpathy This man speaks truth!

@karpathy Actually Warcraft 3 was the peak

StarCraft 2 and Half Life 2 were created perfect, and gaming has been downhill since those times. Is this universally agreed on or just me getting old youtu.be/M_XwzBMTJaM

@ThePhDPlace The guitar! Really academia is an elaborate backup plan to have an interesting day job

Do you have hobbies outside of academia?

New preprint! “Collective behavior from surprise minimization,” by me, @BerenMillidge, @lancelotdacosta, @richardpmann, Karl Friston, @icouzin Paper: arxiv.org/abs/2307.14804 Code: github.com/conorheins/col… 🧵 (1/14)

It was a genuine pleasure and a great honour to discuss the free energy principle, philosophy, and the approach to artificial intelligence that we’re developing at @VERSESAI @helloVERSES with the awesome folks at @MLStreetTalk, one of my absolute favourite podcasts!

@onisillos @ThePhDPlace Shots fired!

New preprint :) Predictive coding networks allow us to perform Bayesian inference on continuous state variables. In this work, we go beyond Bayesian inference, and show how to perform interventional and counterfactual inference.

@lars_sandved Come on over to the dark side..... Julia is great! On a more serious note, part I shows how to recover the classical EFE planning algorithm using Synthetic AIF. The key bit is that EFE as we know it becomes the energy term of a node on the graph. So it's just like evaluating VFE

@KoudahlM Amazing. Really impressive work 👏 I still don’t quite understand what it means to compute EFE without a rollout. But this is making me want to learn Julia properly.

The final project of my PhD is out at last! A comprehensive version of active inference as pure message passing on graphs😎 You want freeform generative models? With epistemics? That are real fast? Direct policy inference instead of rollouts you say? We got em all right here

@lars_sandved Absolutely! The repo with all experiments for both papers should be going public soon, so there's something to dig into if you want to get a head start. It'll definitely migrate into RxInfer proper in the future :)

@KoudahlM Great work! Will you make this available to play with through rxinfer?