Vincent Herrmann
@idivinci
PhD Student at IDSIA with @SchmidhuberAI. Also still a musician.
Since 1990, we have worked on artificial curiosity & measuring „interestingness.“ Our new ICML paper uses "Prediction of Hidden Units" loss to quantify in-context computational complexity in sequence models. It can tell boring from interesting tasks and predict correct reasoning.
Excited to share our new ICML paper, with co-authors @robert_csordas and @SchmidhuberAI! How can we tell if an LLM is actually "thinking" versus just spitting out memorized or trivial text? Can we detect when a model is doing anything interesting? (Thread below👇)
Have you ever wondered what happens when you reduce a story to a low-dimensional latent representation? Well my new IEEE TPAMI work with @idivinci, Zachary Friggstad, and @SchmidhuberAI shows that it has some pretty cool applications 🎉 Check it out: doi.org/10.1109/TPAMI.… (1/n)
Had a great time presenting at #ICML2024 alongside @idivinci & @LouisKirschAI. But the true highlight was @SchmidhuberAI himself capturing our moments on camera! #AI #DeepLearning
Pleasant surprise of @icmlconf: There is a growing Weight Space Learning Community out there - it was great to meet you all: Haggai Maron (@HaggaiMaron), Gal Chechik (@GalChechik), Konstantin Schürholt (@k_schuerholt), Eliahu Horwitz (@EliahuHorwitz), Derek Lim (@dereklim_lzh),…
To celebrate @DavidDeutschOxf’s 70th birthday, colleagues and friends of David have contributed text and videos in which they share ideas on physics and philosophy related to David’s work: dd70th.weebly.com
lol, draft of full GPT-4 paper with architecture and data details is already leaked on torrent😂 The vision component in the architecture is an interesting twist to plain ViT, and scaled up quite a bit! Link to the torrent for the curious: assets.speakcdn.com/assets/2703/bu…
New paper from IDSIA motivated by building an artificial scientist with World Models! A key idea is to get controller C to generate pure thought experiments in form of *weight matrices* of RNNs that still surprise the world model M, and then updating it. arxiv.org/abs/2212.14374
We address the two important things in science: (A) Finding answers to given questions, and (B) Coming up with good questions. Learning one abstract bit at a time through self-invented (thought) experiments encoded as neural networks arxiv.org/abs/2212.14374