Researchers from the National Research Council Canada experimented with four large vision-language models to assess racial and gender bias. They found biases in the models’ evaluation of scenarios in images based on race and gender. Their experiments used a dataset called PAIRS and revealed biases in occupation scenarios and social status evaluations, raising the need…

The advent of large language models (LLMs) has transformed natural language processing, but their high computational demand hinders real-world deployment. A study explores the viability of smaller LLMs, finding that compact models like FLAN-T5 can match or surpass larger LLMs’ performance in meeting summarization tasks. This breakthrough offers a cost-effective NLP solution with promising implications.

Google, led by CEO Sundar Pichai, is shifting focus towards AI chatbot technology with Gemini. This innovative tool aims to offer a versatile and interactive way of accessing information, including text, voice, and images. Google is experimenting with various formats for Gemini and plans to offer advanced features through a subscription model, reflecting a strategic…

This week’s AI news covers a range of topics, including AI’s involvement in defense applications and its impact on carbon emissions. Efforts to combat AI-generated fake content are also discussed, along with developments in AI image generation and its application in different industries. The post concludes with a selection of engaging AI stories.

Physics-informed neural networks (PINNs) integrate physical laws into learning, promising predictive accuracy. However, their performance declines due to multi-layer perceptron complexities. Physics-informed machine learning efforts are ongoing, but PirateNets, designed by a research team, offer a dynamic framework to overcome PINN challenges. It integrates random Fourier features and shows superior performance in addressing complex problems…

Stanford researchers have introduced RAPTOR, a tree-based retrieval system that enhances large language models with contextual information. RAPTOR utilizes a hierarchical tree structure to synthesize information from diverse sections of retrieval corpora, and it outperforms traditional methods in various question-answering tasks, demonstrating its potential for advancing language model capabilities. [47 words]

Large Language Models (LLMs) have become crucial for Natural Language Processing (NLP) tasks. However, the lack of openness in model development, particularly the pretraining data composition, hinders transparency and scientific advancement. To address this, a team of researchers has released Dolma, a large English corpus with three trillion tokens, and a data curation toolkit to…

The Frontiers of General Artificial Intelligence Technology Exhibition in Beijing unveiled a virtual robot toddler named Tong Tong, developed by the Beijing Institute for General Artificial Intelligence. Tong Tong exhibits human-like abilities and behaviors, mirroring those of a 3-4 year old child. Chinese researchers aim to create thousands of powerful autonomous robots by 2025.

MIT researchers have revealed how utilizing symmetry in datasets can reduce data needed for training models. They employed Weyl’s law, a century-old mathematical insight, to simplify data input into neural networks. This breakthrough has potential implications in computational chemistry and cosmology, and it was presented at the December 2023 Neural Information Processing Systems conference.