Researchers from Stanford University, UMass Amherst, and UT Austin have developed a novel family of RLHF algorithms called Contrastive Preference Learning (CPL). CPL uses a regret-based model of preferences, which provides more accurate information on the best course of action. CPL has three advantages over previous methods: it scales well, is completely off-policy, and enables…

Researchers challenge the belief that Vision Transformers (ViTs) outperform Convolutional Neural Networks (ConvNets) with large datasets. They introduce NFNet, a ConvNet architecture pre-trained on the JFT-4B dataset. NFNet performs comparably to ViTs, showing that computational resources are crucial for model performance. The study encourages fair evaluation of different architectures considering performance and computational requirements.

Language models like GPT-3 can generate text based on learned patterns but are neutral and don’t have inherent sentiments or emotions. However, biased training data can result in biased outputs. Sentiment analysis can be challenging with ambiguous or sarcastic text. Misuse can have real-world consequences, so responsible AI usage is important. Researchers at UC Santa…

LLMTime is a method proposed by researchers from CMU and NYU for zero-shot time series forecasting using large language models (LLMs). By encoding time series as text and leveraging pretrained LLMs, LLMTIME achieves high performance without the need for specialized knowledge or extensive training. The technique outperforms purpose-built time series models across various issues and…

ULTRA is a model for learning universal and transferable graph representations for knowledge graphs. It can generalize to any KG with different entity and relation vocabularies, and it outperforms specialized baselines in link prediction experiments. ULTRA’s performance is enhanced through pre-training and fine-tuning, making it suitable for inductive and transferable KG reasoning. Future work includes…

President Joe Biden has issued a comprehensive executive order on AI governance aimed at ensuring transparency and standardization in the industry. The order emphasizes the need for clear content labeling and watermarking practices and includes requirements for AI developers to share safety test results with the US government. Critics have noted the lack of enforcement…

Researchers from Apple and Carnegie Mellon University have developed a benchmark called TIC-DataComp to train foundation models like OpenAI’s CLIP models continuously. They found that starting training at the most recent checkpoint and replaying historical data delivers performance on par with an Oracle while being 2.7 times more computationally efficient. The findings highlight the need…

Despite some progress in the SAG-AFTRA strike negotiations, unresolved issues remain, including the use of AI in recreating performers’ likeness and revenue sharing with streaming platforms. The strike has continued for 109 days, with uncertainty surrounding its end date. Negotiations between SAG-AFTRA and industry producers are ongoing. The Writer’s Guild of America has already secured…

ChatGPT has shown impressive performance in various disciplines, but it struggles with math. While it has performed well in exams like medical and law schools, it falls short in accounting. A study conducted by Professor David Wood revealed that ChatGPT scored 47.4% on accounting exams, significantly lower than humans’ average score of 76.7%. LLMs like…

Amazon Bedrock is a managed service by AWS that provides access to foundation models (FMs) and tools for customization. It allows developers to build generative AI applications using FMs through an API, without infrastructure management. To ensure data privacy, customers can establish a private connection between their VPC and Amazon Bedrock using VPC endpoints powered…