This Week in AI: AlphaFold, Building an Interactive Research Poster, DeepSpeed Compression, and more

This week was a big one for AI in life science with AlphaFold predicting 200 million+ 3D protein structures, allowing us to dramatically increase our understanding of biology. Big tech and universities were busy open sourcing libraries for federated learning, graph neural networks, and optimizing deep learning. And finally, Wikipedia should now become more accurate thanks to Meta's new AI model. Let’s get into it.

Research Highlights

• DeepMind’s AlphaFold expanded their protein database by 200x, from nearly 1 million to 200+ million, predicting nearly every protein known to science. In just 12 months, AlphaFold has been accessed by more than 500,000 researchers and used to accelerate progress on important real-world problems ranging from plastic pollution to antibiotic resistance. See more below or Read more on the DeepMind Blog

"As someone who worked on protein structure modeling in my Ph.D. studies, I was deeply impressed (no pun intended) by the AlphaFold2 approach. AlphaFold outperformed 98 alternative approaches at the CASP13 protein-folding competition and correctly predicted 25 out of 43 protein structures. However, while the database of 200 million structures is impressive, we should bear in mind that the accurate prediction of protein structures is still an active area of research." -- Sebastian Raschka

• Presenting at EMNLP, CCVPR, or another conference? Quickly build an interactive Research Poster to highlight your work.
• In order to accelerate the adoption of ML in structural engineering, Huu-tai Thai at the University of Melbourne published a comprehensive review of the applications of machine learning in structural engineering focusing on advancements in “computational capabilities as well as the availability of large datasets.”
• Researchers in China released the first two datasets for large-scale Small Object Detection (SOD), which aims to increase the performance of computer vision applications like drone scene analysis and traffic sign detection.
• A new dissertation “focuses on improving the robustness and privacy of distributed learning algorithms” to better understand the vulnerabilities of devices and appliances connected to the internet. (Tian Liu, Auburn University)
• Liqun Huang from the Beijing Institute of Technology found that a Transformer-based GAN for Brain Tumor Segmentation performed better than state-of-the-art methods when analyzing image segmentation from MRI scans.

ML Engineering Highlights

• Microsoft announced DeepSpeed Compression, “a composable library that combines novel compression technologies and highly efficient system optimizations to make DL model size smaller and inference speed faster, all with much lowered compression cost.”
• Meta hopes to increase the accuracy of Wikipedia with SIDE, a new AI Model that automatically scans hundreds of thousands of citations and checks whether they support corresponding claims.
• NVIDIA released an overview of evaluating data lakes and warehouses, as well as what's best for ML workloads, finding that both are useful approaches to tame big data and make steps forward for advanced ML analytics.

Open Source Highlights

"If you are a PyTorch user with GNN-FOMO, check out PyTorch Geometric." Sebastian Raschka

• Researchers at the University of Michigan released FedScale, “a federated learning benchmarking suite with realistic datasets and a scalable runtime to enable reproducible FL research.”
• Google open-sourced the alpha release of a scalable Python library for Graph Neural Networks in TensorFlow to encourage GNN productization and experimentation.
• Google introduced LocoProp, a new framework for training DNN models that outperforms comparable higher-order optimizers like Shampoo and K-FAC, “without the high memory and computation requirements.”

Community Spotlight

Want your work featured? Contact us on Slack or email us at [email protected]

• Justin Goheen’s high-level framework for students and researchers needing to incorporate deep learning into a project: https://github.com/JustinGoheen/lightning-pod
• Unofficial implementation of NeRF (Neural Radiance Fields) using Lightning by @kwea123: https://github.com/kwea123/nerf_pl
• A clean and scalable template to kickstart your deep learning project by ashleve: https://github.com/ashleve/lightning-hydra-template

Conferences

• Ai4 2022: AI and Machine Learning Conference August 16-18, 2022 (Las Vegas, Nevada)
• NLP 2022: 11th International Conference on Natural Language Processing September 17-18, 2022 (Copenhagen, Denmark)
• GTC 2022: Developer Conference for the Era of AI September 19-22, 2022 (San Jose, California)
• MLNLP 2022: 3rd International Conference on Machine Learning Techniques and NLP September 24-25, 2022 (Toronto, Canada)

Don’t Miss the Submission Deadline

• BIOS 2022: 8th International Conference on Bioinformatics & Biosciences October 29-30, 2022 (Vienna, Austria) - Submission Deadline August 6!
• ODSC West 2022: Open Data Science Conference November 1-3, 2022 (San Francisco, California) - Speaker Proposal Deadline August 29!
• CCVPR 2022: 5th International Joint Conference on Computer Vision and Pattern Recognition December 9-11, 2022 (Kuala Lumpur，Malaysia) - Abstracts Due September 9!

Upcoming Community Events

• In-Person Panel: Productivity and Performance Boosters for Data Scientists and Developers August 9, 2022
• Livestream: Build an Interactive Research Poster with Lightning AI August 12, 2022
• Webinar: Reduce Infrastructure Cost by Moving Beyond MLOps August 18, 2022
• Meetup: Speed up your Machine Learning Applications with Lightning AI at PyData Miami / Machine Learning Meetup September 22, 2022

Want to learn more from Lightning AI? “Subscribe” to make sure you don’t miss the latest flashes of inspiration, news, tutorials, educational courses, and other AI-driven resources from around the industry. Thanks for reading!