WHAT WAS IMPORTANT
📍 The 35th AAAI conference took place from February 2 to 9 and received a record 9034 submissions compared to the previous year with 8800. The AAAI 2021 Best Paper Awards went to the following teams:
- Haoyi Zhou et al., for applying the Transormer architecture for Long Sequence Time-Series Forecasting in their “Informer” paper (arxiv.org).
- Stefanos Leonardos et al. for Exploration-Exploitation in Multi-Agent Learning: Catastrophe Theory Meets Game Theory (arxiv.org).
- Ruibo Liu et al. who explain how political bias in language models can be mitigated (Paper and Keynote on dartmouth.edu).
🇩🇪 At the end of December, the German Federal Government issued its National AI Strategy updated (bmbf.de) and promised new investments of 5 billion euros by 2025. national data strategy (bundesregierung.de).
While she continues to address privacy and personal data, she is primarily focused on how research can be improved and how to strengthen collaboration with companies and other stakeholders. A remarkable change in an otherwise more technophobic, privacy-focused government (More on netzpolitik.org).
9 THINGS WE FOUND WORTH SHARING
📄 1. Research — The research team behind Andrew Brock from Google's DeepMind presents in the paper “High-Performance Large-Scale Image Recognition Without Normalization” Normalizer-free networks (NFNets) before. By omitting batch normalization, they drastically speed up model training and achieve 89.2% Top 1 accuracy on Imagenet classification at 8.7x faster training as the comparable Efficientnet-B7 (arxiv.org).
Yannic Kilcher, the One-Man Paper Discussion Group, goes into further detail in a recent video (34 min on YouTube.com).
Image: EA/Bioware
💡 2nd use case — Game developer BioWare used AI-powered texture upscalingto remaster the Mass Effect series. The game, released in 2007, was dressed up for current consoles such as PlayStation 5 and massive time savings were possible thanks to AI (venturebeat.com).
💻 3. Libraries & Code — Lightning Flash is a high-level framework for rapid prototyping of deep learning tasks based on PyTorch Lightning. It aims to simplify fine-tuning and the use of cutting-edge models for customer-specific tasks. In this medium article, the developers go into detail (medium.com/pytorch) and the code is available on GitHub (github.com).
🎨 4th showcase — Jina released version 1.0 of her Open-source neural search engine for companies and developers. yours multimodal The search framework can easily find text, graphics, audio, or video in large cloud structures, ideally suited for large amounts of corporate data. Jina.ai makes the code freely available! (jina.ai & github.com).
💭 4.2 Articles & Tutorials — Or for the programmers among you: Start from scratch and build your own AI-powered search engine with Elasticsearch, Kubeflow and Katib following this Topbots tutorial (topbots.com).
👁 5. Miscellaneous — Training facial recognition systems requires large amounts of input images. Some model trainers took the liberty of scraping images from websites like Flickr — and Your picture could be one of them. With Exposing.ai, two tech artists have built a platform to verify whether a specific photo was used to train well-known face recognition systems (exposing.ai).
🛠 6th tools — Paper with Code Just became Papers + Code + Datasets. The popular database for code implementation, benchmarks and scientific work now also contains 3000+ data sets, sorted by task area, modalities and usage statistics (paperswithcode.com/datasets).
💡 7. Use Case — Researchers from Tartu, Estonia, have trained a GAN to “artificial genomes” to create. Instead of creating DeepFakes of human faces, they falsify DNA sequences that are 100% synthetic but indistinguishable from human genomes. This could relieve research institutions of complex data protection measures, which they currently have to ensure with real DNA data (thenextweb.com).
🎨 8th showcase — Automated evaluation of research papers with AI: Are you waiting for your work to be checked? With this tool, you can abbreviate the process. Although only English is currently available, the summaries and reviews are surprisingly well written. (review.nlpedia.ai).
Image: Le et al.
📄 9. Research — Dutch researchers Le et al. have tried to reconstruct Dr. Who scenes from the viewer's fMRI brain scans. “Brain2Pix: Fully Convolutional Naturalistic Video Reconstruction from Brain Activity“(biorxiv.org).
ICYMI
📍 That Enterprise AI Consulting AMAI team has a new website Launched. An updated career page is accompanied by a newly launched AI Expert Developer Blog.
Visit us on www.am.ai
UPCOMING EVENTS
📅 February 24 (online, 11:00 a.m. CET) — Boosting high energy physics with generative networks: In this data science seminar Anja Butter from the Institute for Theoretical Physics in Heidelberg shows how it uses generative neural networks for simulations for the CERN Large Hadron Collider — meetup.com.
📅 March 3 (online) — Robert Bosch AI CON 2021 — Watch the speaker and panel discussion Bosch-ai.com.
