A new open-source system called Dobb-E can train robots for domestic tasks using real home data, addressing the lack of training data in robotics. Utilizing an iPhone and reacher-grabber stick to collect data, the system achieved an 81% success rate in executing household tasks over 30 days. The team aims to expand Dobb-E’s capabilities.

Indiana University researchers have developed Brainoware, a groundbreaking artificial intelligence system that combines lab-grown brain cells with computational circuits to achieve speech recognition and mathematical problem-solving. This innovative technology showcases potential in advancing AI capabilities and poses ethical considerations. While challenges exist, Brainoware offers hope for future applications in neuroscience and computing paradigms.

Image-text alignment models aim to connect visual content and textual information, but aligning them accurately is challenging. Researchers from Tel Aviv University and others developed a new approach to detect and explain misalignments. They introduced ConGen-Feedback, a method to generate contradictions in captions with textual and visual explanations, showing potential to improve NLP and computer…

Google AR & VR and University of Central Florida collaborated on a study to validate VALID, a virtual avatar library comprising 210 fully rigged avatars representing seven races. The study, which involved a global participant pool, revealed consistent recognition for some races and own-race bias effects. The team highlighted implications for virtual avatar applications and…

The study introduces “Vary,” a method to expand the vision vocabulary in Large Vision-Language Models (LVLMs) for enhanced perception tasks. This method aims to improve fine-grained perception, particularly in document-level OCR and chart understanding. Experimental results demonstrate Vary’s effectiveness, outperforming other LVLMs in certain tasks. For more information, visit the Paper and Project.

Meta’s introduction of Emu as a generative AI for movies signifies a pivotal moment where technology and culture merge. Emu promises to revolutionize access to information and entertainment, offering unprecedented personalization. However, the potential drawbacks of oversimplification and reinforcement of biases call for a vigilant and balanced approach to utilizing this powerful tool.

LLM360 is a groundbreaking initiative promoting comprehensive open-sourcing of Large Language Models. It releases two 7B parameter LLMs, AMBER and CRYSTALCODER, with full training code, data, model checkpoints, and analyses. The project aims to enhance transparency and reproducibility in the field by making the entire LLM training process openly available to the community.

Numerical simulations used for climate policy face limitations in accurately representing cloud physics and heavy precipitation due to computational constraints. Integrating machine learning (ML) can potentially enhance climate simulations by effectively modeling small-scale physics. Challenges include obtaining sufficient training data and addressing code complexity. ClimSim, a comprehensive dataset, aims to bridge this gap by facilitating…

The MIT-Pillar AI Collective has selected three fellows for fall 2023. They are pursuing research in AI, machine learning, and data science, with the goal of commercializing their innovations. The Fellows include Alexander Andonian, Daniel Magley, and Madhumitha Ravichandra, each working on innovative projects in their respective fields as part of the program’s mission to…

LLMLingua is a novel compression technique launched by Microsoft AI to address challenges in processing lengthy prompts for Large Language Models (LLMs). It leverages strategies like dynamic budget control, token-level iterative compression, and instruction tuning-based approach to significantly reduce prompt sizes, proving to be both effective and affordable for LLM applications. For more details, refer…