Artificial intelligence, or AI, is an ever-evolving field. Don't be deceived by the occasional failures here and there: its advance is unstoppable, and also causes quite a few worries (I talked about it here and also here).
However, these should not prevent us from also publicizing the great progress that AI has made. In the last year and a half, despite the global crisis (in some cases, indeed, "thanks to" the global crisis), scientists, researchers and developers have achieved unprecedented goals. I don't know if there will be sensational updates between now and the end of the year, but since 2022 is near it seems right to me to already take stock. In the past year, these are the innovations that I believe are milestones in the field of artificial intelligence.
Facebook SEER
Earlier this year Facebook AI (or should I say Meta?) He developed SEER (SElf-supervised). It is a self-supervised computer vision model with billions of parameters. The model can learn from any random group of images on the Internet, without having to carefully curate and label the images (this was a prerequisite for computer vision training). So far, the Facebook AI team has tested SEER on a billion images from Instagram. It is among the milestones because it paves the way for flexible, accurate and adaptable computer vision models for the future.
Isomorphic Labs
This novelty has arrived fresh fresh, I have recently talked about it in this post. Alphabet plans to develop a computational platform to better understand biological systems and find ways to treat disease. Although separate, the two AI systems DeepMind and isomorphic they intend to collaborate occasionally to develop research, discoveries and work on the structure of proteins. It is among the milestones because it will lead to much faster research and development of drugs for all types of pathology.
MusicBERT
Microsoft He developed a large-scale pre-trained model for understanding symbolic music. Is called MusicBERT, and can understand music from symbolic data, i.e. in MIDI format. The tech giant used a method called OctupleMIDI to train its system on a database of more than a million songs. It is among the milestones because today it achieves cutting-edge performance in music understanding tasks and, going forward, it can be trained on tasks that include structure analysis and chord recognition. And one day it will revolutionize music.
GitHub Copilot
The artificial intelligence society OpenAI and Microsoft teamed up to launch the programmer AI Copilot in July this year. Based on OpenAI Codex, the new artificial intelligence system is trained on open source code, and is able to "understand" computer code, generating new strings relevant to the context. It runs on a broad set of frameworks and languages including TypeScript, Ruby, Java, Go, and Python. He is among the milestones because (guess what?) today he is a valid assistant to programmers, tomorrow he will be able to do it himself and revolutionize programming.
Tensorflow 3D
Google developed and launched TensorFlow 3D, a modular library for bringing 3D deep learning capabilities to TensorFlow, earlier this year. This latest update gives access to sets of operations, loss functions, models for developing, training and deploying 3D scene understanding models, and data processing tools and metrics. It is among the milestones because it will help robots and autonomous vehicles to perfectly recognize the 3D objects that come across their path (as well as improving virtual reality environments).
