2022 Annual Review | More Mature AI, More Disruptive Technology Carnival
2022 is coming to an end amidst the coughing sounds. This year, AIGC has become the biggest dark horse in the field of artificial intelligence, ScienceAI has more practical applications, this year we have experienced budget cuts and layoffs by major manufacturers, and also experienced a technological carnival of survival... This article,Let’s review with our colleagues the groundbreaking research and development achievements in the field of artificial intelligence in 2022.
data2vec
A general framework for self-supervised learning of speech, vision, and text
Publishing Agency:Meta AI
Release time:January 2022
Project address:
data2vec is aA unified multimodal self-supervised learning model,It can process images, text, voice and other tasks with higher performance.
On December 16, data2vec 2.0 was released, which is 16 times faster than existing computer vision self-supervised algorithms with the same accuracy.
AlphaCode
Competition-level code generation AI
Publishing Agency:DeepMind
Release time:February 2022
Project address:
AlphaCode uses a large language model to construct code based on the language description of the problem.In the Codeforces challenge, AlphaCode defeated 46%. The research not only appeared on the cover of Science, but was also named one of the top ten scientific breakthroughs of the year by the magazine.
Dall E 2
Text to image generation tool
Publishing Agency:OpenAI
Release time:April 2022
Project address: https://openai.com/dall-e-2/
Dall·E 2 Based on the text description, you can create artistic images with more realistic colors.Compared to Dall·E released by OpenAI in 2021, Dall·E 2 can generate more realistic and accurate images with 4 times the resolution.
Gato
All-round intelligent agent
Publishing Agency:DeepMind
Release time:May 2022
Project address:
Gato is an all-round intelligent agent.It can play Atari games, describe images, chat, and output text, joint torque, or other tokens based on the context.
This general model solves all tasks similar to artificial intelligence and may eventually surpass domain-specific models.
ESM Fold
Protein structure prediction models
Publishing Agency:Meta AI
Release time:July 2022
Project address:
ESM Fold is a model for predicting protein sequences.It can directly perform highly accurate, end-to-end, atomic-level structure prediction. It uses only a single input sequence and only needs to look at a single protein sequence, which greatly speeds up inference.
Make-A-Video
AI system that generates videos from text
Publishing Agency:MetaAI
Release time:September 2022
Project address: https://makeavideo.studio/
Make-A-Video is a text-to-video generation model that learns common descriptions from images with text descriptions, and understands and learns movement patterns using unlabeled videos.
The videos generated by Make-A-Video have diverse styles and high text restoration, making it a SOTA model for generating short videos.
AlphaTensor
Improve matrix multiplication and increase calculation speed
Publishing Agency:DeepMind
Release time:October 2022
Project address:
AlphaTensor improves the current optimal 4*4 matrix multiplication and further improves the calculation speed of more than 70 other matrix multiplications of different sizes. This achievement was featured on the cover of Nature.Named one of the top ten scientific breakthroughs of the year by Scinece magazine.
Magic 3D
Text-to-3D content creation tools
Publishing Agency:NVIDIA
Release time:November 2022
Project address:
NVIDIA joins AIGC3D Mesh models can be generated based on text descriptions.It combines image conditioning technology and text-based prompt editing methods to provide a new way to control 3D synthesis, making it possible to create high-quality 3D Mesh models.
ChatGPT
Super Conversation Model
Publishing Agency:OpenAI
Release time:November 2022
Project address:
ChatGPT is trained using RLHF (Reinforcement Learning from Human Feedback), the same method used by InstructGPT, with only slight differences in the data collection setting.
ChatGPT can chat and communicate like a human, and complete tasks such as writing emails, video scripts, copywriting, translation, and coding. Since its launch, it has attracted countless developers at home and abroad to try it out and have heated discussions. It can be said to be the technical project with the highest developer participation in 2022.
Point E
Generate 3D point cloud based on text description
Publishing Agency:OpenAI
Release time:December 2022
Project address:
The process of generating a 3D point cloud based on text prompts using Point·E is divided into three steps::
1. Generate a synthetic view based on the text prompt
2. Generate a coarse point cloud (1024 points) based on the synthetic view
3. Generate fine point cloud (4096 Point) based on low-resolution point cloud and synthetic view
Using a single Point·E card, 3D point cloud can be generated in 1 minute, and text-to-3D says goodbye to the era of high computing power consumption.
Winter is gone and spring is here, let’s imagine 2023
2022 is coming to an end, and 2023 is destined to be a year full of unknowns. What new achievements will there be in the field of AIGC? How will ScienceAI respond to the challenges brought about by the intersection of basic science and AI? What new breakthroughs will be made in chip research and development and domestic operating systems?
What are your predictions for the technologies and applications in the field of artificial intelligence in 2023? Welcome to leave a message to discuss~
Chao Neuro also has many articles introducing the development of artificial intelligence in the past year. Click to read~