mixing is a wrong metaphor IMHO because they are only per token same with Microsoft-China (aka ResNet) connections. The only mixing is attention. And obviously all frontier LLMs hack this a lot, especially for agentic rollouts. Same with profile/sampling separation.

Stepfun在努力，恭喜 张祥雨老师和孙剑老师的论文 ResNet 获得CVPR2026 Longuet-Higgins Prize （时间检验奖）！ 阶跃 Step 3.7 Flash 拿下 Artificial Analysis 多个第一！ 搭车招人！！ PS: 我不是男娘，我是研发工程师，不是hr/运营！！！！！！welcome to talk！！！

Building AI-assisted ultrasound interpretation for under-resourced clinics in Kenya, Kenya has roughly 1 radiologist per 270,000 people, so most rural areas have zero coverage. We've got a working biometry model (ResNet-18 on HC18 data), backend mobile app in progress.

Great read. Yes the zero padding of the 3x3 conv in ResNet (and ConvNeXt etc) is actually what makes a convnet LESS good to use in a sliding manner than a ViT, which sounds crazy at first!

Noci et al. explicitly cited Stable Resnet paper and acknowledged the overlap, whereas this work did not. That said, both papers are interesting contributions.

I don't think it's a grand conspiracy or anything, just that we seem to have created an odd ahistorical narrative where we're a bit bored of summer internationals altogether and resnet having to go play them. I resent the bloody lions, to be frank, don't care about 'em...

データが少なく、リアルタイム性が求められる画像認識プロジェクトで選定に迷っている企業担当者へ ResNetは残差学習で深層化を安定化させ、主要フレームワークやエッジチップで最適化済み。独自データが10万件未満や推論10ms以下の現場では最も高いROIを実現する傾向がある。 最新モデルへ移行すべきか、ResNetを極限まで最適化して運用すべきか、どちらが事業に合理的だろうか？ai-market.jp/technology/what…

最新モデルに移行すべきか悩む経営・開発担当へ ResNetは2026年でも推論速度とROIの最適解。独自ドメインで学習データが10万件未満、または推論レイテンシが10ms未満の現場では特に有利で、エッジ最適化も成熟済み。 まずはデータ量とレイテンシ要件を基準に検討してみては？ai-market.jp/technology/what…

训练神经ODE这种连续深度模型，理论分析常是黑盒——我们不知道学习动力学到底怎么演化。不理解这点，就难设计可控训练策略，也没法把结论推广到ResNet、自回归等常见架构。这篇论文从统计物理引入动态平均场理论（DMFT），在高维极限下把多体相互作用简化为单粒子过程，通过自洽方程精确求解在线SGD的学习曲线，首次为神经ODE提供了从前向传播到训练误差的解析刻画。 arxiv.org/abs/2606.07247

Day 12 -Worked on CIFAR-10 in PyTorch today. Added batch normalization and data augmentation to my CNN, improved its accuracy to ~81%. Also tried a pretrained ResNet-18 and learned more about transfer learning. #Day12 #ADAS #Pytorch #ComputerVision

1/5: Robustness to Input Corruptions We evaluate pretrained models on corrupted vision and language data: ImageNet-C and FineWeb10B-C. Across ResNet, ViT, GPT-2, and GPT-2 Medium, Muon achieves the best overall robustness, suggesting that Muon learns more corruption-resistant features.

■DenseNetとは？ 前に見つけた特徴をみんなで使い回すCNN ✅ResNet 前の情報を一部ショートカットして後ろへ渡す ✅ DenseNet 過去の情報をどんどん共有。そのため、画像の特徴をムダなく使えて、 少ない計算でも学習しやすくなる #AI #G検定 #今日の積み上げ_892

Is the task simply harder in extreme regimes? Partly — but that's not the explanation. A ResNet-18 trained from scratch on raw pixels outperforms every frozen VFM probe in exactly this regime. The representations lost the physical resolution present in the pixels. And the geometry confirms it: in intense regimes, the features collapse onto each other — distinct storms become nearly indistinguishable in latent space.

zkCompute Hub : Where Heavy Computing Meets Hard Money in Web3. Imagine you’re working on an AI project that requires hundreds of hours of GPU time. Or you need to generate complex ZK proofs. Or you want to render high-quality 3D animations. Until now, your options have been limited: send it to an expensive, centralised cloud provider, or entrust it to random people on the internet, hoping they won’t cheat and that the results will be correct. This is where zkCompute Hub comes in. This isn’t just any ordinary marketplace. It’s a decentralised, verifiable compute marketplace built on @LitecoinVM — Litecoin’s Virtual Machine. A place where anyone can post computational tasks, anyone can work on them, and most importantly: the results can be cryptographically verified before any money changes hands. The Real Problem We Face Heavy-duty computing is a necessity for the future. AI training, ZK proof generation, scientific simulation, rendering, and even large-scale data processing — all require expensive, centralised computing power. But there are two major issues: 1. Trust Issue. How can you be sure the work you’ve paid for has actually been done correctly? What if the results are manipulated? 2. Slow and Complex Settlement. Even if the results are correct, the payment process is often manual, rife with negotiation, and prone to disputes. zkCompute Hub eliminates both these issues with one simple principle: don’t trust, verify. How does zkCompute Hub work? The process is designed to be very simple, yet powerful: 1. A Poster creates a job (e.g., “Train a ResNet-50 model with dataset X for 50 epochs”) and places the reward in escrow via a smart contract (in zkLTC or USDC). 2. A Worker views the job on the Marketplace and claims it if they feel capable. 3. The worker performs the computation on their own hardware (which could be a home GPU, a server, or a cluster). 4. Upon completion, the worker submits a proof — which could be: • A ZK Proof • A hash check • Or a manual review (for specific cases) 5. The smart contract verifies the proof. If valid, the reward is automatically released to the worker. If not, the escrow is returned to the poster. Everything happens on-chain on LitVM. No third party holds the funds. There is no “trust first, pay later”. What Makes zkCompute Different? Many projects claim to offer “decentralised compute”, but few are truly verifiable. zkCompute Hub takes verifiable compute seriously: • Supports a wide range of tasks: AI Training, AI Inference, ZK Proof Generation, 3D Rendering, Scientific Simulation, RAG Pipeline, FHE Computation, Data Labelling, Video Transcoding, and much more. • Features an on-chain Reputation Leaderboard system, so workers who frequently provide valid proofs gain greater trust. All of this runs on LitVM, meaning its security and finality are backed by the Bitcoin ecosystem. It’s not just another “ordinary blockchain”. Who Can Use It? • AI researchers & builders who need large-scale computation without having to trust cloud providers. • ZK teams that frequently generate proofs and want the process to be cheaper and verifiable. • Studios & creators who need large-scale rendering or transcoding. • Anyone wishing to run heavy computations in a decentralised manner with a guarantee of trustworthy results. On the other hand, anyone with computing hardware (GPUs, CPUs, even clusters) can become a worker and receive automatic payment once their proof is accepted. A Greater Vision zkCompute Hub is not just about “running jobs on the blockchain”. It is about building Hard Money Compute — an ecosystem where computing holds the same high standards of trust as the money backing it. Built on LitVM, this project brings the spirit of Bitcoin (sound money, decentralisation, resilience) to the realm of computing. Imagine a day when major AI projects, research laboratories, or even game studios can “hire” computing power from thousands of machines worldwide — without worrying about results being manipulated, and without the need for large intermediaries taking a hefty cut. That is the vision of zkCompute Hub. zkCompute Hub is currently available to try on the LitVM LiteForge Testnet (Chain ID: 4441). Live app: zkcompute-hub.vercel.app Open source: GitHub: github.com/fens21/zkCompute-… Computing is one of the most valuable resources of this century. Yet, until now, it has remained trapped in a centralised world built on blind trust. zkCompute Hub aims to bring it into a world that is more open, fairer, and more verifiable. Because in a world increasingly reliant on computing, trust should no longer be about who is doing the work, but whether the results can be proven. 📍Note This app is still under development and contains numerous bugs and system errors. If you choose to use it, please use a testnet wallet (a wallet that holds no monetary value whatsoever). #hackathon_submision #hackathon_litvm @circle_crypto

産総研政策予算・画像基盤モデルチームの全体像です。FDSL, Ego4D, 3D ResNet, LIMIT などの参画メンバーが集い、視覚/3D/動画/マルチモーダル基盤モデルの構築を効率的に実施しつつ、知見を横展開しながらアップデートします。 共同研究や連携ラボから、各種応用も見据えた実装を進めています。

最近したこと KV260くんにResnetとかEfficientNet-B0とかYOLOとかのCNNを構築して検証しまくるお仕事

■WideResNet 深さよりも幅を広げて効率よく学習できるようにしたResNet ✅ResNet スキップ結合を使って深い層でも学習しやすくしたが、層を深くすると計算が大変になったり学習効率が悪くなる ✅WideResNet 層の数を増やしすぎず1つ1つの層で扱う情報量を増やす #AI #G検定 #今日の積み上げ_891