bugsinside @buginsideout

This is the best book I will ever read . Extremely satisfying

Alex Bilzerian @alexbilz

2 years ago

'Classical Electrodynamics' - J.D. Jackson (PDF): fisica.ugto.mx/~ggutj/CV/Clas…

3 36 253 39K 185

In a #ICML2026 position paper we argue a dynamical systems perspective is needed to drive time series models forward: arxiv.org/abs/2602.16864 We need to move away from transformers that do not respect a system’s dynamical structure, esp. for out-of-domain generalization & insight

Jensen-Shannon centroid of normalized histograms: ➊ Set of histograms form a *mixture family* ( information geometry) ➋ Transform JSD centroid into an equivalent Jensen centroid on mixture parameters + applies ConCave-Convex Procedure (CCCP)

MAD: Manifold Attracted Diffusion ift.tt/i6kCp0U

[LG] MiniMax Sparse Attention X Lai, W Xu, Y Yang, Q Chen… [MiniMax] (2026) arxiv.org/abs/2606.13392

keyboard_arrow_left Previous keyboard_arrow_right Next

on this beautiful saturday afternoon we’re going to read about how vector policy optimization inject solutions diversity in these models

Evaluating the Representation Space of Diffusion Models via Self-Supervised Principles Xiao Li, Yixuan Jia, Zekai Zhang, Xiang Li, Lianghe Shi, Jinxin Zhou, Zhihui Zhu, Liyue Shen, Qing Qu arxiv.org/abs/2606.09718 [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙲𝚅] 💬Accepted at ICML 2026

Our TMS algorithm is now published in @ImagingNeurosci doi.org/10.1162/IMAG.a… If you are coming to @OHBM @OHBM_Trainees, come check out our poster about our open label trial. Poster number 1152 Monday, Jun 15 | 14:45-15:45 Tuesday, Jun 16 | 13:30-14:30

Thomas Yeo @bttyeo

2 months ago

We develop a new TMS targeting algorithm and test it in an open label trial in a treatment-resistant depression population with high comorbidities. Preprints by @rubykong92 @Torphernchern 1. doi.org/10.1101/2025.0… 2. doi.org/10.64898/2026.… Our new approach ...

2 39 88 11K 30

Optimizing Energy-based Neural Network Training with Coherent Ising Machine Chen-Rui Fan, Bo Lu, Zhi-Hong Zhang, Run-Qing Zhang, Jing-Wei Wen, Chuan Wang arxiv.org/abs/2606.09117 [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸]

LEAF: A Learning-Enabled ADMM Framework for Accelerated Convex Optimization Binh Nguyen, Trinh Tran, Truong X. Nghiem arxiv.org/abs/2606.08993 [𝚌𝚜.𝙻𝙶 𝚎𝚎𝚜𝚜.𝚂𝚈 𝚖𝚊𝚝𝚑.𝙾𝙲]

A Quadratic Order Reduction -- Gaussian Process Ordinary Differential Equation framework for the inference of Large Continuous Dynamical Systems ift.tt/JojDL4e

Restoration of vision in adult mammals remains unattainable following complete transection of the optic nerve. However, using transplanted human #RetinalOrganoids, researchers at @CQU1929 were able to partially restore vision in rats! 🔬 bit.ly/43iLB8l 🎧 bit.ly/4dYixJs

I am happy to share our work that redefines the vertebrate testis as a highly rhythmic organ, after almost three decades of misunderstanding stemming from its lack of strong daily cellular rhythmicity. Science Advances science.org/doi/10.1126/sc…

Did Anthropic get more gains out of model scaling than other labs thought was possible? It reminds me of an interesting recent paper, which showed that deep layers in open LLMs are not doing much, and that this can be fixed by scaling the LayerNorm output. arxiv.org/abs/2502.05795

CATASTROPHE THEORY FOR SCIENTISTS AND ENGINEERS By Robert Gilmore “This advanced-level treatment describes the mathematics of catastrophe theory and its applications to problems in mathematics, physics, chemistry and the engineering sciences.” — Publisher description (Wiley, 1981 / Dover reprint)

Towards More General Control of Diffusion Models Using Jeffrey Guidance ift.tt/uSkoWXe

This paper took nine years and was, honestly, a labor of love and a fair amount of suffering. It took several labs, sustained NIH support, and a lot of people who stuck with it, a whole village! It is one of the largest cocaine self-administration studies ever done in rats. Definitely my biggest experiment (so far). Most self-administration study runs dozens of animals. We phenotyped 836, with extended access over several weeks, plus measures of escalation of intake, motivation under progressive ratio, and continued use despite footshock. A quick breakdown of the results. First, we replicate our earlier behavioral analysis, which used about half as many animals (doi.org/10.7554/eLife.…): escalation of intake, motivation to seek cocaine, and compulsive-like responding are tightly correlated and cluster together. Now, with more animals we now resolve finer structure. The behaviors separate into early acquisition, late acquisition, escalation, and a motivational/compulsive factor that groups breaking point, responding during timeout, and responding during footshock. That grouping makes a lot of sense. Second, these behaviors are heritable. Phew, good ;-) SNP-based heritability came in around 7 to 16% depending on the trait. That is modest and, as expected, well below twin-study estimates, but it is real. The most heritable phenotype was a composite measure of behavior during escalation after several weeks of self-administration, and 7 of the top 10 most heritable phenotypes came from long access. That validates the importance of the long-access model. Third, we found six loci across five chromosomes tied to different addiction-like measures. One is unusually clean: a narrow region containing a family of carboxylesterase genes, the enzymes that metabolize cocaine, with a direct ortholog to the human enzyme CES1. It was associated with the time between infusions, an interval long proposed as a measure of compulsive-like intake, and the effect size (0.75 SD) is pretty large sufficient to change an individual from mild to severe. Finding a metabolic enzyme echoes human GWAS for cocaine and tobacco use disorder, which also point to pharmacokinetic enzymes for alcohol and nicotine. A nice crossspecies translational convergence on the idea that metabolism genes contributes to addiction vulnerability. This also converges with an existing therapeutic effort. For 15+ years, groups have worked on engineered carboxylesterases to treat cocaine overdose and addiction, without knowing that genetic variants in that same enzyme are actually tied to cocaine vulnerability. That work has moved slowly, partly because pharma has shown little interest. Our data give the direction more reason to exist. We also picked up Trak2, which has been identified in human cocaine use disorder GWAS. Trak2 helps traffic GABA receptors in the brain, and GABA signaling is something we and others have found to be central to drug self-administration. On top of those two genes, we replicated several from human alcohol and tobacco use GWAS, including SLC10A7, PLCL1, and SATB2. Finding these parallels between human and animal data is exciting, especially now, when animal models are often dismissed as useless for understanding mental illness. They are not, and this is a clear example of why. Finally, no GWAS is much fun without new candidates. We found several genes carrying expression or splicing QTLs in relevant brain regions, including Rasd2, Gnas, Ctsz, Lsm6, Vsnl1, and Zfp831, Slc6a2, a number of them tied to dopamine/norepinephrine signaling. Thanks to @kallupi_marsida , @giodeguglielmo , and @LotCarrett in my lab for an enormous amount of behavioral work, to @AbePalmer's group @PsychiatryUcsd and first author @MontanaKayLara for the genetic analysis, and to Leah Solberg Wood @WakeForest for the HS rats. And a real thank you to our @NIDAnews program officers and the study section reviewers who backed this epic study when it was far from a sure thing 8 years ago... doi.org/10.1038/s41467… @NatureComms

keyboard_arrow_left Previous keyboard_arrow_right Next

Great to see the fundamental advantages of nonlinear RNNs! Why parallelizing them during training is so important!

Anamaria-Roberta Hartl ✈️ ICML @anamariarp17

5 days ago

New paper "On Subquadratic Architectures: From Applications to Principles" 🙌 On commonsense & reasoning benchmarks, xLSTM, Mamba-2 & Gated DeltaNet perform nearly indistinguishably. Therefore, we analyse them where structure genuinely matters: on code & time series.👇

3 10 27 6K 21

1/8 We published a new paper from @fastinoAI: Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers. Main idea: adaptive optimizers like AdamW, Sophia, and Shampoo do not just have noisy updates. Their stochastic preconditioned updates are biased in specific, correctable ways. Paper: arxiv.org/abs/2605.20756.

Can we improve energy-based diffusion models by solving a classification task? 🤔 🚀 Excited to share DiffCLF, accepted at ICML 2026 🇰🇷! ! We predict a sample’s noise level, and recover densities as a by-product. Joint work with @theh2o64 (co-lead) and @jmhernandez233. [🧵1/n]