🚀 What is Trending in AI Research?: SoTaNa + PointHPS + Lemur project + VirSen1.0 + Jais and Jais-chat...

➡️ How can we create a more effective and accessible software development assistant, given the limitations of existing large language models like ChatGPT and LLaMA in understanding human intent and limited accessibility?

This paper introduces SoTaNa, an open-source assistant for software development. SoTaNa leverages ChatGPT to generate high-quality, instruction-based data in the software engineering domain and fine-tunes LLaMA for enhanced performance. The model is evaluated using Stack Overflow questions and also assessed for code summarization and generation capabilities. Importantly, SoTaNa is designed to run on a single GPU, expanding its accessibility.

➡️  How can accurate Human Pose and Shape Estimation (HPS) be achieved from 3D point clouds, especially when these data are noisy, incomplete, and feature diverse human poses?

Researchers at NTU Singapore propose PointHPS, a principled framework designed to handle these challenges. PointHPS uses an iterative, cascaded architecture to refine point features from 3D point clouds captured in real-world settings. The framework employs downsampling and upsampling at each stage to glean both local and global cues. Two innovative modules are introduced: 1) Cross-stage Feature Fusion (CFF) for efficient multi-scale feature propagation, and 2) Intermediate Feature Enhancement (IFE) for body-aware feature aggregation, thereby improving the feature quality. Tests conducted on two large-scale benchmarks demonstrate that PointHPS significantly outperforms existing state-of-the-art methods.

➡️  XLang NLP Lab Researchers Propose Lemur: The State-of-the-Art Open Pretrained Large Language Models Balancing Text and Code Capabilities

The Lemur project, spearheaded by XLang Lab in collaboration with Salesforce Research, seeks to address this critical gap in language model technology. Lemur and Lemur-chat represent a pioneering effort to develop open, pretrained, and supervised fine-tuned LLMs that excel in both text and code-related tasks. The cornerstone of this endeavor is the extensive pretraining of Llama 2 on a vast corpus of ~100 billion lines of code-intensive data. This pre-training phase is followed by supervised fine-tuning on ~300,000 instances of public instructional and dialog data. The result is a language model with enhanced coding and grounding abilities while retaining competitive textual reasoning and knowledge performance.

➡️ What's the most efficient way to design a sensor-based human gesture recognition system, especially when current methods suffer from poor interactive performance and lack of useful tools?

Researchers from the University in Yokohama propose VirSen1.0, a virtual environment equipped with a user interface designed to streamline the development process. The environment houses a simulator that generates "lightness data" and combines it with an avatar's motion to train machine learning classifiers. One of the standout features is the interface's ability to visualize the importance of individual features used in the learning model via permutation feature importance. This offers developers real-time feedback on the efficacy of each sensor, thereby aiding in quicker iterations. A user study confirmed that this feature visualization helps in achieving models with better accuracy.

➡️ Researchers from Inception, MBZUAI, and Cerebras Open-Sourced ‘Jais’: The World’s Most Advanced Arabic Large Language Model

This paper introduces Jais and Jais-chat, large language models built on the GPT-3 architecture, but optimized for Arabic. With 13 billion parameters, these models are pretrained on a dataset comprising both Arabic and English texts, as well as source code in various programming languages. Their performance outstrips existing open Arabic and multilingual models in knowledge and reasoning tasks by a significant margin. Interestingly, they also perform competitively in English, despite being trained on less English data. The paper delves into the details of training, tuning, safety alignment, and evaluation, and offers two open versions of the models to encourage further research in Arabic language models.

What is Trending in AI Tools?

Editor’s Recommended AI Tool

Bubble: Bubble lets you create interactive, multi-user apps for desktop and mobile web browsers, including all the features you need to build a site like Facebook or Airbnb. Build out logic and manage a database with our intuitive, fully customizable platform. Get started for free.