AI News about Automated Image Descriptions and more

1. Code – Uber open sourced Kepler.gl its geospatial analysis tool for large-scale data sets. Kepler is a web-based application built on top of Mapbox GL and deck.gl that can render millions of points on the fly (github.com).

2. Education – Graph Neural Networks (GNNs) are likely to make further advancements in 2021. To understand the basics watch this 8-minute introduction from AI Coffee Break (youtube.com).

3. Paper – The authors from Heidelberg University use Transformers to generate high-resolution images in "Taming Transformers for High-Resolution Image Synthesis" (arxiv.org).
👉 Try it yourself by creating landscape images with this Colab notebook (colab.research.google.com).

4. Use Case – VFX artists from Corridor Digital compare AI-powered techniques with traditional workflows in visual effects. They use Runway ML (runwayml.com) to drastically cut time on tasks such as rotoscoping, inpainting and camera-tracking (youtube.com).

5. Code & Tutorials – LabML published a collection of Neural Net PyTorch implementations with side-by-side formulas and notes explaining each step. The implementations include recent publications such as Google's 1.6 trillion parameters Switch Transformer (labml.ai).

6. Paper – The authors of "ZeRO-Offload: Democratizing Billion-Scale Model Training" introduce a deep learning training method that allows training of multi-billion parameter models on a single GPU (arxiv.org).

7. Article – Kamales Lardi presents lessons for company leaders in his article "Understanding The Value Of Artificial Intelligence Solutions In Your Business" (forbes.com).

8. Use Case – In 2016 Facebook introduced automated alternative text (AAT) to aid blind or visually impaired users. The system describes image content and aims to make FB's platforms more inclusive. Despite recurring privacy concerns (cnet.com), they utilize facial recognition to identify contacts in images.
The latest version of its AAT includes information about the relative size of objects and positional context, such as what is in the center of the image or on something else. Furthermore, the system can identify monuments and landscape thus allowing the visually impaired to understand image context better. – Read more on the models and training methods used in the Facebook Tech Blog (tech.fb.com).