This Month in AI

CVPR

🎸 Winning an honorable mention at CVPR, CMU researchers developed a method to sense high-speed visual vibration with two low-speed vibration cameras. Their method can reconstruct sound from multiple simultaneous vibrations, such as instruments in a band or a packet of crisps in a room while someone’s talking.

🪞 The Ref-NeRF paper also won an honorable mention for enabling Neural Radiance Fields (NeRFs) to generate more realistic transparent and glossy surfaces. Instead of the original view-dependent outgoing radiance, they introduce a reflected radiance.

👩 ETH Zürich and Max Planck Institute researchers developed IMavatar, which produces a 3D avatar of an individual’s head from a monocular video.

🏙️ ContrastMask outperforms Mask-RCNN and other state-of-the-art architectures in instance segmentation. It’s designed to improve discrimination between foreground and background features.

🏃 Researchers from the Technical University of Munich, NVIDIA, and CALTECH propose an active learning framework that considers a model’s robustness using a novel inconsistency score.

At V7

🎓 We went to CVPR! See how our MNIST socks made it to CVPR’s Twitter :)

We also launched our Free Education Plan and invited first users to the platform. Are you a student, researcher, or a professor working on an interesting computer vision use case?

You are in for a treat! Apply for a free access to V7 by filling in the form.

Industry

🔎 Google released an open-source dataset of photo-realistic 3D object models. The dataset contains one thousand household items ready for interactive simulations and model training.

💻 Trained on billions of lines of code, GitHub Copilot is now available to the public. The AI-assisted programming tool can make suggestions and automatically complete sections of code.

⚡ Google Cloud partners with French energy company ENGIE to predict the energy output of wind farms thirty-two hours in advance. As a result, ENGIE will be able to minimize dependence on fossil fuels.

Research papers

🤏 Microsoft researchers propose XTC, a compression method that uses extreme quantization and layer reduction to reduce the model size of transformers like BERT by up to 50x.

🚀 DeepNorm stabilizes thousand-layer transformers. Proposed by Microsoft researchers, it prevents deep models from getting stuck during training by scaling down residual connections.

⛏️ OpenAI released a model that can play Minecraft. Using Video PreTraining on more than 70,000 hours of gameplay, OpenAI’s agent can collect materials and craft wooden and stone tools.

📈 Google, Stanford, and UNC researchers discuss the sharp performance jumps in language models after they reach a certain size.

👀 We are hiring...

1️⃣ Applied Researcher (Computer Vision)

2️⃣ Deep Learning Engineer (Internship - Computer Vision)

3️⃣ Technical Lead, Deep Learning (Computer Vision)

See you in July! 👋