Artificial Intelligence
Researchers have created a virtual reality app that allows users to open and control 3D modeling tools simply by moving their hand.
In the context of voice assistant systems, steering is when users issue follow-up commands to direct or clarify previous turns. We propose STEER, a model that detects whether a follow-up turn is the user attempting to steer the previous command. To train the model, we created heuristic rules to sample usage data without annotations, overcoming…
EELBERT is an approach for compressing transformer-based models like BERT while preserving accuracy in downstream tasks. It replaces the input embedding layer with dynamic embedding computations, reducing model size. Evaluations on the GLUE benchmark demonstrate the effectiveness of this approach.
Amazon CodeWhisperer is a generative AI coding companion that helps developers optimize their code for sustainability. It provides recommendations for code improvement based on existing code and natural language comments, allowing developers to reduce resource usage and energy consumption. CodeWhisperer supports multiple programming languages and is available as part of various development tools. Unoptimized code…
QB-GPT is a model that can generate football plays based on provided elements. It aims to recreate plays from minimal information to understand how player setups and contextual elements affect team paths on the field. The model uses data from the Next Gen Stats (NGS) tool of the NFL and relies on the GPT architecture.…
The development of MetNet-3 represents a significant breakthrough in meteorological research, addressing challenges in weather forecasting. This comprehensive neural network model integrates various data sources, such as radar data and satellite images, to generate precise and high-resolution weather predictions. With advanced techniques like topographical embeddings and model parallelism, MetNet-3 has the potential to revolutionize the…
Steak-umm’s latest advertising campaign involved confronting a group of vegans with deepfake videos of them enjoying meat. While the vegans initially reacted angrily, they eventually supported the campaign’s goal of highlighting the impact of deepfakes. Steak-umm hopes the campaign will attract meat eaters and also raises awareness for the proposed Deep Fakes Accountability Act, which…
Machine learning (ML) is increasingly important across industries, but there is a gap between business expectations and what engineers and data scientists can deliver. The first step to close this gap is fostering honest dialogue between teams and then democratizing ML across the organization. Companies can enable collaboration between technical and non-technical teams, provide the…
Samsung Electronics held the Samsung AI Forum 2023 to discuss generative AI and its impact on daily life and work. Samsung Research introduced its generative AI model, Samsung Gauss, highlighting the company’s commitment to this technology. Industry and academic leaders shared insights on large language models, multimodal AI technology, and graduate students presented groundbreaking research.…
The UK AI Safety Summit and Biden’s executive order have brought AI regulation into focus, but questions remain about the specifics. The Bletchley Declaration, endorsed by 28 countries, emphasizes international consensus on AI oversight. The US and EU have proposed their own regulations, while other countries consider their own initiatives. The implementation of regulations across…
ChipNeMo explores the use of domain adaptation techniques to improve the performance of language models (LLMs) in chip design. The study evaluates three LLM applications in chip design and highlights the potential for further refinement in domain-adapted LLM approaches. The goal is to enhance LLM performance and reduce model size while maintaining or improving performance…
Atom is a new low-bit quantisation technique developed by researchers to increase the serving throughput of Large Language Models (LLMs). By using low-bit operators and quantisation, Atom reduces memory usage without sacrificing precision, resulting in improved end-to-end throughput by up to 7.73 times compared to existing approaches. Atom addresses the need for more efficient LLM…
US export restrictions on Nvidia have created a growing market in China for Huawei’s new AI chips, specifically the Ascend 910B. Chinese AI companies are turning to Huawei’s chip as a viable alternative to Nvidia’s high-end chips. The export controls, intended to slow Chinese AI innovation, may have inadvertently accelerated China’s path to self-reliance. As…
This article discusses various methods to style plots using Matplotlib. It covers topics such as changing runtime configuration parameters, creating and using style files, applying style sheets, and limiting styling to code blocks. These techniques allow for customization and consistency in plotting styles.
Chinese researchers have developed a deep learning model called circ2CBA that can predict binding sites between circular RNAs and RNA-binding proteins. This has significant implications for understanding diseases, particularly cancer. The model uses sequence information and a unique process to accurately identify these critical interactions, surpassing existing methods. The results validate the effectiveness of circ2CBA…
Researchers at the University of Oxford have introduced DynPoint, an artificial intelligence algorithm that enables the rapid synthesis of novel views for unconstrained monocular videos. DynPoint employs explicit estimation of consistent depth and scene flow for surface points, creating a hierarchical neural point cloud to generate views of the target frame. The proposed model demonstrates…
This research paper introduces a method called “codebook features” that aims to enhance the interpretability and control of neural networks. By leveraging vector quantization, the method transforms the dense and continuous computations of neural networks into a more interpretable form by discretizing the network’s hidden states. The experiments conducted demonstrate the effectiveness of codebook features…
Researchers from the University of Tokyo have developed a deep learning model called 3D-Memory In Memory (3D-MIM) to accurately predict the expansion of supernova (SN) shells in galaxy simulations. By combining the model with the Hamiltonian splitting method, the researchers can integrate SN-affected particles separately. The 3D-MIM model shows strong generalization capabilities and offers a…
Big language models (LLMs) are becoming skilled in programming and refactoring code to create libraries for software developers. Researchers from MIT CSAIL, MIT Brain and Cognitive Sciences, and Harvey Mudd College present LILO, a neurosymbolic framework that integrates LLMs with automatic refactoring to learn libraries of reusable function abstractions. LILO demonstrates improved performance compared to…
The ControlLLM framework, developed by researchers from The Hong Kong University of Science and Technology, OpenGVLab, Shanghai AI Laboratory, Tsinghua University, and SenseTime, enables large language models (LLMs) to utilize multi-modal tools for solving complex real-world tasks. ControlLLM excels in accuracy, efficiency, and versatility, surpassing existing methods in various tasks involving image, audio, and video…