8 THINGS WE FOUND WORTH SHARING
🎨 1st showcase - A team of researchers from ETH Zurich and Google spoke at the NeurIPS conference last year Hi-fic presented, short for High-Fidelity Generative Image Compression. They use generative adversarial networks (GANs) to create a State of the art Method for Lossy image compression to create. The amazing results can be seen below. One interactive demo is available online (hific.github.io).
Pictures from: Mentzer et al.
📄 2. Research - Elon Musk's Neuralink is working on a method to convert brain activity into text. For example, this could help paralyzed people communicate via writing on a display. Yours Brain-computer interface (BCI) can generate 90 characters per minute!
Dr. Zsolnai-Fehér, who is behind the YouTube channel Two Minute Papers Is written, appears in Neuralink's BCI paper “High-Performance brain-to-text Communication via imagined handwriting” a (paper on biorxiv.org & the Two Minute Papers summary on YouTube.com).
💡 3rd use case - Measure the sky with AI: gravitational lensing are astronomical objects that bend the path of light. This influences the image that we can perceive from our earthly perspective. These gravitational lenses can be difficult to detect. That's why researchers have trained a Deep ResNet to help them find these artifacts in space images (phys.org)
📰 4th article - Srinivasa Ramanujan (1887 - 1920) was an Indian mathematician who was famous for making important contributions to this field without any formal education in mathematics (he was portrayed in the book and film “The Man Who Knew Infinity”). His notebooks were filled with theorems that turned out to be true but were only proven decades later.
Now AI comes into play: The Ramanujan Machine wants to emulate its namesake by automates new math formulas discovered. However, as with the original Ramanujan, proving the discovered theorems will be left to his successors - (Read more at nature.com).
👁 5. Miscellaneous - Safari is the last major web browser (except Firefox) that, by default, Voice recognition function added (cnet.com). Soon, more websites will add their own voice commands to their online presences, following what Google has already started (9to5google.com).
💭 6th tutorial - Quoc Le and Kip Kähler from Roblox describe how they BERT-Accelerated inference by more than 30x to serve more than 1 billion requests a day with low latency (Robloxtechblog.com).
💻 7. Code & Tutorial - The powerful wav2vec2 Facebook AI is the current state of the art in Speech-to-text transcription. It was recently added to the Huggingface Transformers library (huggingface.co) added.
The following tutorial takes an audio clip from the movie The Dark Knight and converts it to text with this Colab notebook (10 min on YouTube.com & colab.research.google.com).
huggingface.co
💭 8th article - Facebook has its automatic Hate speech detection improved. With their RIO and the self-supervised Linformer They now recognize the model (both published) 97% of hateful posts automatically before a person marks them (tech.fb.com).
UPCOMING EVENTS
💙 March 3 (online, German, 16:00) — State of AI — AI application ideas in logistics — Our managing director Dr. Jürgen Stumpp presents robust AI applications for the supply chain and invites discussion with members of the BVL Federal Logistics Association. — Registration bvl.de.
📅 March 3 (online) — Robert Bosch AI CON 2021 — 300 AI experts from industry and science meet. Speakers from Carnegie Mellon University, the Max Planck Institute and ETH Zurich are invited. — Register here at Bosch-ai.com.
