Today on the blog, learn how we developed simple and effective fixes to the dual challenges of spurious features and simplicity bias by applying early readouts and feature forgetting, resulting in substantially improved generalization to unseen domains. ↓
RESEARCH
-
Research Advances in Text Embeddings Language Models and Prompting
By
–
a tiny summary of my research projects of 2023: Text Embeddings Reveal (Almost) As Much As Text EMNLP 2023, http://
arxiv.org/abs/2310.06816 Language Model Inversion ICLR 2024, http://
arxiv.org/abs/2311.13647 TreePrompting EMNLP 2023, http://
arxiv.org/abs/2310.14034 -
AI Containment Essential for Realizing Technology Benefits
By
–
"I believe these technologies can deliver an extraordinary amount of good. But without what I call containment, every other aspect of a technology, every discussion of its ethical shortcomings, or the benefits it could bring, is inconsequential."
-
Document Search Retrieval RAG Open Source Guide
By
–
Document Search and Retrieval using RAG Great guide by @LightningAI and @aniketmaurya on building a retrieval pipeline Uses open source embedding and reranking models! https://
lightning.ai/lightning-ai/s
tudios/document-search-and-retrieval-using-rag?view=public§ion=recent
… -
NLP and ML for Mental Health Crisis Detection in Clinical Workflows
By
–
Using NLP to detect mental health crises: “Data science and ML can be successfully integrated into clinician workflows, leading to dramatic improvements when it comes to identification of patients at risk,” says @Stanford student @akshay_swa
. -
Time-LLM: Forecasting Time Series with Large Language Models
By
–
Time-LLM: Time Series Forecasting by Reprogramming Large Language Models Jin et al.: https://
arxiv.org/abs/2310.01728 #ArtificialIntelligence #DeepLearning #MachineLearning -
BPE algorithm’s longest tokens are banners and whitespace
By
–
It’s whatever the byte-pair encoding algorithm says which depends on the pre-train statistics. There are many shorter ones too — the top 130 longest tokens are all banners and whitespace of various lengths.
-
Runway ML Mitigates Stereotypical Biases in Text-to-Image Systems
By
–
Our new research on mitigating stereotypical biases in text to image generative systems has proven to significantly improve fairness across different groups of people.
— Runway (@runwayml) 2 février 2024
Learn more about our diversity fine tuning (DFT) approach: https://t.co/zseBIVAlHs pic.twitter.com/xv91Qg5MgjOur new research on mitigating stereotypical biases in text to image generative systems has proven to significantly improve fairness across different groups of people. Learn more about our diversity fine tuning (DFT) approach: https://
research.runwayml.com/publications/m
itigating-stereotypical-biases-in-text-to-image-generative-systems
… -
Media amplification of controversial AI research ideas
By
–
Le plus inquiétant n’est pas tant qu’un chercheur ponde une idée pareille, mais plutôt qu’un grand quotidien comme Libé se fasse l’écho de ces élucubrations, contribuant ainsi à la fabrique de l’opinion.
-
SVD vs Stable Video: Video Generation Model Comparison
By
–
"Tilt down" camera effect (which I've looped). SVD remains my favourite video model, it is the least "melty".
— fofr (@fofrAI) 2 février 2024
No interpolation controls in Stable Video at the moment, but it looks like some are being applied to the output to smooth it out. https://t.co/kdP3hfYSad pic.twitter.com/XiJ6ERxKOl"Tilt down" camera effect (which I've looped). SVD remains my favourite video model, it is the least "melty". No interpolation controls in Stable at the moment, but it looks like some are being applied to the output to smooth it out.