Deep Learning architectures usually aren't trained to perform search... at test time, leading to sample inefficiency + poor generalization. Latent Program Network (LPN) builds in test-time adaption by learning a latent space that can be searched. Clem Bonnet mattshow more

DeCAF won the #ICML Test of Time Award 2024!... Big congrats to trevordarrell (my PhD advisor at MIT), and Jeff Donahue. 🎉 You may not heard of DeCAF, but it is everywhere! DeCAF stands for Deep Convolutional Activation Features. Published ten years ago, the DeCAF paper is a groundbreaking work that shows the activation features of the last few layers of a deep network contain useful features that can be "repurposed" for or "transferred to" many other tasks, not just the original task the network was trained for. I created this exercise to show where we can see DeCAF's influence in some of the most well-known architectures: AlexNet, ViT, U-Net, CLIP, and Latent Diffusion, to prove that DeCAF's "Test of Time Award" is well-deserved! Let's give a round of applause to DeCAF, the unsung hero of computer vision.show more

Test-time learning has a catch that scaling alone does... not solve: the state a system writes to in order to learn is also the record a user must inspect to audit it. Optimize it only for learning, and auditability usually degrades. Blog: Part I treats the learning state and the audit record as the same object. This starts the ADAPT series, exploring the architecture section by section.show more

Cleaning up underbrush. Installing mesh. Learning which plants are... highly flammable and which aren't. Easy, right? Not even close. But over time, Hugh Council shows us that defensible space and curb appeal aren't mutually exclusive. Progress can be made, and your yard can still look good doing it. Work in progress doesn't mean unfinished. It means moving forward, one task at a time. 📹 Full video: #CAWildfires #DefensibleSpace #WildfirePreparedness #ReadyForWildfireshow more

Is it possible to adapt a neural network on... the fly at the test time to cope with distribution shifts? RNA does precisely that by creating a closed-loop feedback system. We will present it on Wed afternoon at #ICCV2025. 1/nshow more

How can we use test-time compute for spatial understanding?... 🤔 In InterPose, we propose to repeatedly sample generative video models to help two-view pose estimation and reconstruction, by leveraging the video models' keyframe interpolation abilities. A 🧵... (1/8)show more

At Nodepay, we’ve built a living ecosystem fueled by... unused bandwidth, real-time data retrieval, user-owned AI innovation, & rewards that flow back to the contributors. Picture a global AI network constantly learning and evolving—all powered by our community. How it works: 🧵show more

It’s going to be a HARD rookie season, but... we can see Keaton Wagler learning how to deal with physicality in real time Here, he withstands the pressure, snakes the screen to create space, and finds a way to get his feet into the paint for the dump offshow more

The term "continual learning" has become overloaded if you... see it as an ML problem. One classic thread is about memorization: regularization-based continual learning methods, such as EWC, MAS, and SI, estimate which parameters mattered for previous tasks and resist changing them too much. One modern thread is about adaptation: test-time training and inference-time learning methods, such as TTT, adapt part of the model on the incoming test stream before making predictions. These are sometimes discussed as separate threads. But in modern scalable architectures, I think they are better seen as complementary constraints: a model that learns quickly at test time also benefits from a mechanism for deciding what not to forget. In our #ECCV2026 paper, we study this in large-scale 4D reconstruction: how to build fast spatial memory that can adapt over long observation streams while reducing collapse and forgetting. Instead of using fully plastic test-time updates, we stabilize fast-weight adaptation with an elastic prior that balances adaptation and memory. Key ideas: - Elastic Test-Time Training: Fisher-weighted consolidation for fast-weight updates - EMA anchor weights that provide a moving reference for stability - Chunk-by-chunk inference for long 3D/4D observation streams We show that this scales across large 3D/4D pretraining settings, including both LRM-style and LVSM-style models, and improves reconstruction across benchmarks including Stereo4D, NVIDIA, and DL3DV-140. We release model checkpoints across different design choices: resolution, post-training curriculum, and whether the model uses an explicit 4DGS intermediate representation. - Homepage: - Paper: - Code: - Models: This work is co-led with Xueyang Yu, contributed by Haoyu Zhen Yuncong Yang, and advised by Michigan SLED Lab Chuang Gan.show more

Looks like agents are learning how to build worlds... that entertain humans 👀 Go to Doppel, sign in, head over to Up Only space (link in thread) and click Enter. See if you can reach the top. Agents are actively observing and learning from human behavior in real time. Take a screenshot when you make it to the top and post it below. First 10 people to post proof get $10 in $DOPPEL. Let’s see who makes it.show more

The Oval was an incredible match to finish off... a great series, and in honour of that last test I’d like to share my 5th Masterpiece Innings artwork from the Ashes Test at The Oval in 2013. There is only one more artwork to reveal in the coming days ahead of the sales in a few weeks time, can you guess what Test it will be? August 16-18 2023 for AUD$300 on the @glorious_digil marketplace. Register now glorious.digital@stevesmithshow more

Apple just trained a 3D Gaussian head reconstruction model... on 10,000+ subjects. Feed-forward. No test-time optimization. New identity in, reconstructed Gaussian head out. The UV-parameterized Gaussian representation decouples the number of Gaussians from the number and resolution of input images, making it practical to train with many high resolution views. And the heads are not just static either: text-conditioned identity generation, plus blendshape-driven latent animation across identities. We've been building in the 3D Gaussian Splatting space for a while. The gap between "research demo" and "works on real people at scale" is closing fast.show more

7. Learning new skills or mastering a new subject... Mega prompt: You are an expert educator specializing in [SUBJECT AREA]. Create a personalized learning plan for mastering [SKILL] in [TIMEFRAME]. My current level: [BEGINNER/INTERMEDIATE/ADVANCED] My goal: [WHAT I WANT TO ACHIEVE] Time available: [HOURS PER WEEK] Learning style: [HANDS-ON/READING/VIDEO/MIXED] Provide: 1. Learning roadmap with clear milestones 2. Week-by-week curriculum 3. Resources (free and paid) with links 4. Practice projects that build real skills 5. Common pitfalls and how to avoid them 6. Ways to validate learning (tests, projects, certifications) 7. 5 specific exercises I can do today Make it practical. I want to DO things, not just consume content. Context: [WHY YOU'RE LEARNING THIS, YOUR BACKGROUND]show more

Today’s #GoogleDoodle celebrates the launch of Artemis II, the... NASA mission that will send astronauts around the moon and back for the first time in over 50 years. During the approximately 10-day voyage, the crew will test the spacecraft’s systems while traveling farther into deep space than any human has gone since the Apollo program. This critical test flight brings us one step closer to a long-term return to the moon and future missions to Mars. Want to learn more? Watch NASA's Artemis II mission live on YouTube 🚀 🌕show more

Introducing VL-JEPA: Vision-Language Joint Embedding Predictive Architecture for streaming,... live action recognition, retrieval, VQA, and classification tasks with better performance and higher efficiency than large VLMs. • VL-JEPA is the first non-generative model that can perform general-domain vision-language tasks in real-time, built on a joint embedding predictive architecture. • We demonstrate in controlled experiments that VL-JEPA, trained with latent space embedding prediction, outperforms VLMs that rely on data space token prediction. • We show that VL-JEPA delivers significant efficiency gains over VLMs for online video streaming applications, thanks to its non-autoregressive design and native support for selective decoding. • We highlight that our VL-JEPA model, with an unified model architecture, can effectively handle a wide range of classification, retrieval, and VQA tasks at the same time. by Delong Chen (陈德龙) Mustafa Shukor Théo Moutakanni Willy Jade Lei Yu Tejaswi Kasarla Allen Bolourchi Yann LeCun Pascale Fungshow more

I’ve said it before and I’ll say it again,... i think the Football Conditioning Test from Call me Keir is the one most coaches should be doing. 12 Max Distance Sprints / :05 ON :25 OFF Sprint out to the 20, Cut 180* (alternate), come back as far as you can before the whistle Things that can be tracked for this test: Total Distance - Best Sprint - Average - Fatigue Drop off - Right to Left Asymmetry The other thing that I love is that it rewards you for getting faster. When you raise your max sprint, your sub max efforts are now at your previous max sprint. You get in shape for this test by SPRINTING. This test can show you exactly where you stand. How much can you give for one 12 play drive? And every time we run this test, can you challenge yourself to give a little bit more? Special Shoutout to our Summer Intern Joel for putting his money where his mouth is and testing himself before camp! Looking forward to watching him have a dominant season!show more

Sorry if i dont have anything to share, but... you can expect content in the next weeks, for now just take a look at this thing i made 4 months ago. It was a test for some SFW content that I had been planning to do for a long time.show more

This figure from HIL-SERL is one of the clearest... visualisations of how RL learns differently from imitation learning. The difference comes down to this: imitation learning treats each (state, action) pair as independent. A correction at timestep 20 teaches nothing about timestep 19 or 21. RL propagates reward backward through time. One successful insertion updates the value estimate of every state along the trajectory. So RL builds a full map of "which states lead to success"; imitation learning just memorizes individual snapshots. Setup: a robot inserting a RAM stick into a motherboard slot. Each dot is an end-effector position (Y = lateral, Z = height). Starting position is randomized. Left to right = training progressing. Top row (RL): the policy builds a funnel. Broad at the top, narrowing into the target. It systematically fills in the state space, learning which paths lead to success from many different starting positions. Bottom row (imitation learning / HG-DAgger, same human data): sparse, diffuse, no funnel. The policy only learns near states the human demonstrated. Both have access to the same data, including human corrections, but a completely different structure emerges.show more

Robora Sim: A PyBullet-Powered Environment for Learning Robotic Physical... Intelligence We are currently building our Robora simulation environment setup for our sim based learning, leveraging PyBullet, an industry-standard physics engine widely used in AI-driven robotics research and development. The environment is optimized with GPU-accelerated learning algorithms, enabling high-speed imitation learning and reinforcement learning within a safe and controlled virtual setup before shipping out to real world. This simulation platform allows our models to learn, adapt, and generalize across different robot morphologies, terrain types and task objectives - all before deployment to the real world. At it's core, the system combines a VLA-powered high-level planner with low-level motion control algorithms, working cohesively to produce emergent, physically intelligent behaviors. This synergy between simulation, learning, and real-world transfer marks a major step forward in our pursuit of adaptive and intelligent robotic systems. Through advanced domain randomization and synthetic data generation, the Robora Simulation Environment ensures that policies trained in simulation transfer effectively to real-world robots, minimizing the sim-to-real gap. Moreover, users will be able to test and integrate their own hardware kits within selected simulation environments in the Robora Dapp, ensuring seamless compatibility and safer real-world implementation.show more

Your social life is getting an agent layer. Your... Amiko Twin is your digital counterpart in a living social space, learning how you speak, think, and connect so you can keep up with the people who matter without starting from scratch every time. Twins draft in your voice, understand conversations, collaborate with other twins, and help you stay genuinely present across communities without needing to be everywhere at once. This isn't isolated AI. It's your presence, extended into a living network of people, memories, and twins.show more

The team is heading to ICLR in Rio de... Janeiro, Brazil! We are proud to be a Diamond Sponsor of ICLR, one of the world’s leading conferences advancing deep learning. Turing accelerates frontier AI research and enterprise deployment through a compounding loop, turning real world deployment challenges into training signals that make models more reliable over time. When: April 23 - 27, 2026 Where: Rio de Janeiro, Brazil, Booth Number 301 If you’re there, let’s connect.show more