AI Roundup 03/02 -> 03/09/2023

Last week's most interesting AI news in <2 minutes

1. Scientists fronm Osaka University used Stable Diffusion models to read brain scans and re-create largely realistic versions of images participants seen at ~80% accuracy (Paper's website)

2. Researchers from Google, UC Berkeley, and University of Michigan present Preference Transformer - a neural architecture that models human preferences using transformers, eliminating the reliance on human feedback (Paper)

3. Microsoft researchers introduce ITG (Interactive Text Generation), a new model to train text, image, and code generation models without the costs of involving real users, outperforming other Transformer-based models (Paper)

4. Researchers from NVIDIA and the Imperial College London introduce Prismer, an open-source vision-language model that performs reasoning tasks such as image captioning and visual VQA with SOTA performance with substantially less training data (Paper's website)

5. Microsoft researchers propose FoundationTTS, a text-to-speach model that is capable of synthesizing high-quality speech that closely matches the naturalness and similarity of human speech (Paper)

6. Google researchers unveil a new version of its 2B parameters USM (Universal Speech Model), supporting 1,000 languages, and outperforming OpenAI's Whisper for all segments of ASR (Automatic Speech Recognition) (Google Blog)

7. Google researchers propose PaLM-E, a 562B parameters multimodal AI that uses visual data to enhance its language processing capabilities to control robots in the real world (Paper's website)

8. Microsoft presents Visual ChatGPT - a system for interacting with images, using Visual Foundation Models and ChatGPT (Paper's Github repo)

   

9. Google Robotics a new framework for robot control called Grounded Decoding that combines language models and grounded model objectives (Paper's website)

   

11. A former YC founder introduces SLAAP (Self-learning Agent for Performing APIs), that searches, learns and creates API calls, with an automatic retry mechanism upon wrong information retrival (Twitter)

12. Stability AI releases Pick a Pic, an open-source that aims to improve text-to-image generation by collecting data to help align models with human preferences (Twitter)

13. Meta AI releases MuAViC (Multilingual Audio-Visual Corpus) - the first benchmark that makes it possible to use audio-visual learning for highly accurate speech translation (Meta AI)

Last week’s AI Tidbits