Changmin Yu (@changminYu_) header
Changmin Yu (@changminYu_) profile picture

I study computational neuroscience.

London, England
Joined March 2019
Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)Photo by Changmin Yu (@changminYu_)
"Learning Theory from First Principles" (MIT Press, 2024) is a rigorous, 496-page graduate-level textbook exploring the mathematical foundations of machine learning, from statistical learning theory and optimization to neural networks, generalization, and overparameterized models. Freely available, it's an invaluable resource worth adding to your digital library. di.ens.fr/~fbach/ltfp_book.p…
14
202
5
1,502
53,085
Changmin Yu retweeted
Ne dis pas à ChatGPT « analyse ça ». Ne dis pas à ChatGPT « fais-moi un plan ». Ne dis pas à ChatGPT « donne-moi des idées ». Tu utilises une IA ultra-avancée comme si c'était un stagiaire. Voici 10 prompts à copier-coller :
211
4,194
91
21,334
3,164,679
Finally published: “Symmetry as a common origin of hierarchical emergence and abstraction” science.org/doi/10.1126/scia… Symmetry generate hierarchical macroscales and shape the structure of Bayesian beliefs, making inference of high-dimensional systems tractable
22
143
13
857
38,361
What can behavior tell us about neural dynamics? Review by Fanny Cazettes (@FannyCazettes) & Marieke L. Schölvinck tinyurl.com/2dw3675k
47
1
186
7,205
Changmin Yu retweeted
Gemini 4 Carbon is FREE for 6 months 😳 researchers and scientists get 10,000 slots what you get: - Google Antigravity's highest tier - Gemini 4 Carbon - Gemini 4 Argon applications are open, rolling selections.. no rush needed but don't sleep apply here: docs.google.com/forms/d/e/1F…
54
134
6
1,372
275,117
How does the brain keep the body in balance? I introduce an HMM-based control framework and test it against leading theories of brain function: control-as-inference, the free energy principle, variational inference, and MOP. Finally out! mdpi.com/1099-4300/28/11/123…
1
7
21
1,362
Changmin Yu retweeted
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over. And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams. So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super happy to finally release "The Ultimate MoE Training Handbook", free and open, organized around the questions training teams hit: - how many experts, how big, shared or not? - why is half my step spent in all-to-all, and how do I hide it? - why is my run 25x slower with no error in the logs? - softmax, sigmoid or sqrt(softplus) router, and why keep it in fp32? - aux loss or DeepSeek's loss-free bias? - upcycle a dense checkpoint or train from scratch? - can I serve with fewer experts per token than I trained with? ⚠️ Warning: it's long (~70k words, budget a few days, sorry) but with 55 interactive figures, code you can run and numbers from our own runs. Read it here: huggingface.co/spaces/huggin…
140
624
82
4,560
276,886
Changmin Yu retweeted
Andrej Karpathy (ex-OpenAI, ex-Tesla) built a loop that lets AI run ~100 experiments while you sleep, 99k stars on GitHub someone turned its rules into a prompt that makes Opus 5.5 run any job as a graph that actually finishes it's f*cking unreal... send it to Opus 5.5 and thank me later. then read the article below ↓
10
12
1
83
17,438
When we build statistical models of real neurons, they often end up near criticality, suggesting they're finely tuned to the edge of a phase transition. Here we show how criticality emerges naturally (without fine-tuning) in models of large populations 👇 arxiv.org/abs/2609.09438
5
26
3
151
7,571
Changmin Yu retweeted
How well can you control a stochastic system with limited information? Excited to share a new preprint answering this question! “On the Information Required for Feedback Control”, with @jbetancourt015, @EmonetLab, @machtagroup, and @m_abski.
2
24
4
90
14,494
Changmin Yu retweeted
Online now: Recurrent connectivity shapes spatial coding in hippocampal CA3 subregions dlvr.it/TTQFBM
19
76
8,089
Excited to share our #ICML2026 spotlight✨: Hierarchical successor representation for robust transfer. openreview.net/forum?id=txs... We ask a simple question: can RL agents generalise efficiently under changing goals given its existing knowledge about the environment structure? 1/
1
1
149
Empirically, these hierarchical predictive maps support stronger sample efficiency of learning, transfer efficiency to new goals, and stability. These improvements persist even when options are not provided a priori and need to be constructed online. 3/
1
65
Huge thanks to my collaborator and mentor Máté Lengyel. Interested in hearing more? Come talk to me during poster session 3 (9.30am on July 8th at HALL A #203). And as a bonus, I can give you some spoilers on our ongoing work of modelling hippocampal place cell with HSR. 4/4
45
A hierarchical CAN reproduces goal-directed theta sweeps, and predicts further results on goal-dependent phase coding and goal-directed egocentric directional tuning in place cells.
1
2
141
Huge thanks to my amazing collaborators: @zilong_ji , Jake Ormond, John O'Keefe, and especially @NeilBurgess10 for constantly pushing me to move beyond trivial questions and think deeper.
2
127