WHATS IMPORTANT
🧬 DeepMind publishes enormous Protein Database
In what DeepMind themselves call the "most significant contribution AI has made to advancing scientific knowledge", the British research lab launched the AlphaFold Protein Structure Database, which offers the most complete and accurate picture of the human proteome to date. – Read the announcement on deepmind.com
Accompanying, they published a scientific paper (nature.com) and open-sourced the code (github.com) for their impactful protein folding model AlphaFold 2. Just in time, as another research team has achieved similar performance – More on competing model RoseTTAFold on techcrunch.com
👩💻 Facebook AI releases BlenderBot 2.0
- The new chatbot is the first to build long-term memory and search the internet to enrich conversations with users. Moreover, the research lab at Facebook AI has released papers, models, and open-sourced the code. – Find out more at marktechpost.com
Source: Facebook AI Research (FAIR)
- Meanwhile, Facebook's prompt release is drawing criticism. Anders Arpteg, director of Data Science of the Swedish intelligence service, warns on LinkedIn of the system's potential misuse:
8 THINGS WE FOUND WORTH SHARING
🗞 Industry News – TensorRT is an SDK for high-performance Deep Learning inference, allowing apps to perform "up to 40X faster" than CPU-only platforms. NVIDIA's runtime got even faster in the newest version, cutting down inference with the BERT-Large model to unprecedented 1.2 milliseconds. – Full update on nvidia.com
💡 Use Case – In the Karlsruhe Institute of Technology canteen, AI-based image recognition is now used to charge for meals. The autonomous visioncheckout from local company auvisus allows researchers and students to get their food much quicker. – Checkout at auvisus.com
💡 Use Case – In an interview with VentureBeat, Bonnie Titone, CIO at Duke Energ, recalls how the utility provider saved $74 million by employing computer vision and drones to inspect their solar farms. – Full interview on venturebeat.com
🎓 Education – With MadeWithML Engineer Goku Mohandas (goku.me) has collected a grand library of open-source resources to study machine learning fundamentals & MLOps. His project-based collection for software engineers, college grads, and product managers alike has already collected 26,000+ stars on GitHub. – Start learning on madewithml.com
🔬 Research – Last week, the International Conference on Machine Learning conference was held (Learn about ICML and other important AI conferences in our blog post). There, researchers from IBM, MIT, and Harvard released the AGENT dataset, which allows benchmarking an AI model’s "common sense", its notion of objects and agents in an environment. – ICML announcement on ibm.com
🔬 Research – Google presented two connected methods to synthesize high fidelity images, Cascaded Diffusion Models (CDM) and Super-Resolution via Repeated Refinements (SR3). Diffusion models were originally introduced in 2015 (arxiv.org) and have recently experienced a revival due to advantages over other (image) synthesis methods, such as GANs. – More on marktechpost.com
💻 Code & Tools – When handling tabular data, Pandas DataFrames are the typical go-to in Data Science. The alternative Polars DataFrame library outshines Pandas in speed, especially with large-sized datasets. – Find a speed comparison for common DataFrame operations on github.com
🎨 Showcase – Turn yourself into a cartoon for your next Zoom call. With Snap Camera, the company behind Snapchat has brought popular Snapchat filters to the desktop. – The Verge explains how to Pixar-ize yourself
