Artificial Intelligence
OpenAI has revised its usage policies to permit the use of its AI products in certain military applications and is collaborating with the Pentagon on various projects, including cybersecurity and combatting veteran suicide. Although the company previously prohibited military use, the updated terms stress that the tools must not cause harm or be used to…
Meta, led by Mark Zuckerberg, has announced its ambition to develop Artificial General Intelligence (AGI) and plans to make it open-source upon completion. This marks a significant shift for Meta, previously focused on product-specific AI. It aims to combine its AI research groups and invest heavily in infrastructure to achieve this goal. The move raises…
Raesetje Sefala, a South African activist, is using computer vision and satellite imagery to address the effects of spatial apartheid. She aims to map out and analyze racial segregation in housing, hoping to prompt systemic change and equitable resource allocation. Her work is providing valuable data to policymakers and organizations advocating for social justice and…
The Deep Manifold (Variational) Graph Auto-Encoder (DMVGAE/DMGAE) approach by researchers at Zhejiang University presents a method for attributed graph embedding. It addresses the crowding problem and enhances stability and quality of representations by preserving node-to-node geodesic similarity under a predefined distribution, demonstrating effectiveness in extensive experiments. The research aims to facilitate further application through code…
GenCast, a new generative model from Google DeepMind, revolutionizes probabilistic weather forecasting. By utilizing machine learning, GenCast efficiently generates 15-day forecasts with superior accuracy and reliability compared to leading operational forecasts. This advancement marks a significant step in embracing machine learning to enhance weather prediction, with broad implications across various industries and decision-making processes.
Machine learning’s push for personalization is transforming fields such as recommender systems, healthcare, and finance. Yet, regulatory processes limit its application in critical sectors. Technion researchers propose a framework, r-MDPs, and algorithms to streamline approval processes while preserving personalization, showing promise in simulated environments. This work marks a notable advancement in deploying personalized solutions within…
Language models are crucial for text understanding and generation across various fields. Training these models on complex data poses challenges, leading to a new approach called ‘easy-to-hard’ generalization. By initially training on easier data and then testing on hard data, models demonstrate remarkable proficiency, offering an efficient solution to the oversight problem. This approach opens…
In this week’s AI news roundup: – AI creates a comedic show mimicking George Carlin, raising ethical concerns. – CES 2024 highlights AI innovation in products like Samsung Galaxy S24 series and AI For Revenue Summit. – OpenAI’s GPT Store hosts AI “girlfriends” and reciting ChatGPT for poetry. – The rise of deep fake content…
“Puncc, a Python library, integrates conformal prediction algorithms to address the crucial need for uncertainty quantification in machine learning. It transforms point predictions into interval predictions, ensuring rigorous uncertainty estimations and coverage probabilities. With comprehensive documentation and easy installation, Puncc offers a practical solution for enhancing predictive model reliability amid uncertainty.”
The study discusses the challenges in AI systems’ adaptation to diverse environments and the proposed In-Context Risk Minimization (ICRM) algorithm for better domain generalization. ICRM focuses on context-unlabeled examples to improve out-of-distribution performance and emphasizes the importance of context in domain generalization research. It also highlights the trade-offs of in-context learning and advocates for more…
Advancements in Large Language Models (LLMs) enabled by Natural Language Processing and Generation have broad applications. However, their biased representations of human viewpoints stemming from pretraining data composition have prompted researchers to focus on data curation. A recent study introduces the AboutMe dataset to address these biases and the need for sociolinguistic analysis in NLP.
The emergence of large language models has led to rapid advancements in Mixture-of-Experts (MoE) architecture. The DeepSeekMoE model introduced by DeepSeek-AI innovatively addresses challenges in expert specialization through fine-grained expert segmentation and shared expert isolation. Experimental results demonstrate the scalability and performance superiority of DeepSeekMoE, with potential at an unprecedented scale of 145B parameters.
DeepMind’s AlphaGeometry, a new AI system, excels in solving complex Olympiad-level geometry problems, achieving a milestone in AI’s ability for mathematical problem-solving. By combining a neural language model with a symbolic deduction engine and using synthetic training examples, it outperformed previous AI models, approaching human gold medalist levels. This breakthrough opens new possibilities for mathematics…
Advancements in artificial intelligence and machine learning have revolutionized molecular property prediction in drug discovery and design. The SGGRL model from Zhejiang University introduces a multi-modal approach, combining sequence, graph, and geometry data to overcome the limitations of traditional single-modal methods. The model’s intricate fusion layer produces more accurate predictions, marking a potential breakthrough in…
The Billion-Scale Approximate Nearest Neighbor Search Challenge at NeurIPS aims to advance large-scale ANNS. Pinecone’s innovative algorithms excelled across all four tracks: Filter, Sparse, OOD, and Streaming. Pinecone demonstrated exceptional performance, outperforming the winners by up to 2x, solidifying their position as a leader in vector search technology. [49 words]
Recent research from UC Berkeley and New York University explores the deficiencies in multimodal large language models (MLLMs) caused by visual representation issues. The study uncovers the shortcomings of pre-trained vision and language models and introduces a new benchmark, MMVP, to assess the visual capacities of MLLMs. The researchers propose Mixture-of-Features (MoF) methods to enhance…
Scientists have created a soft fluidic switch using an ionic polymer artificial muscle, capable of lifting objects 34 times its weight with ultra-low power. Its small size and light weight allow for use in industrial areas like soft electronics, smart textiles, and biomedical devices, offering precise fluid control in tight spaces.
Ed Newton-Rex, former VP of Audio at Stability AI, has launched ‘Fairly Trained,’ a non-profit certifying generative AI companies for ethical training data practices, aiming to address concerns over data scraping and copyright infringement. The initiative has already certified nine companies and introduced the ‘Licensed Model certification’ to ensure ethical use of training data.
The StableRep model improves AI training by using synthetic imagery to generate diverse images from text prompts, addressing data collection challenges and offering more efficient and cost-effective training options.
The text discusses the potential risks and limitations of relying on external servers for AI applications. It introduces Jan as an open-source alternative that operates entirely offline, addressing privacy concerns. Jan is designed to run on various hardware setups, offering customization and seamless integration with compatible applications. With a commitment to open-source principles, Jan presents…