Pretrained Large Language Models (LLMs) achieve strong performance across a wide range of tasks, yet exhibit substantial variability in the various layers' training quality with respect to specific downstream applications, limiting their downstream performance. It is therefore critical to estimate layer-wise training quality in a manner that accounts for both model architecture and training data. However, existing approaches predominantly rely on model-centric heuristics (such as spectral statistics, outlier detection, or uniform allocation) while overlooking the influence of data. To address these limitations, we propose LayerIF, a data-driven framework that leverages Influence Functions to quantify the training quality of individual layers in a principled and task-sensitive manner. By isolating each layer's gradients and measuring the sensitivity of the validation loss to training examples by computing layer-wise influences, we derive data-driven estimates of layer importance. Notably, our method produces task-specific layer importance estimates for the same LLM, revealing how layers specialize for different test-time evaluation tasks. We demonstrate the utility of our scores by leveraging them for two downstream applications: (a) expert allocation in LoRA-MoE architectures and (b) layer-wise sparsity distribution for LLM pruning. Experiments across multiple LLM architectures demonstrate that our model-agnostic, influence-guided allocation leads to consistent gains in task performance.

Breakthroughs, discoveries, and DIY tips sent every weekday. A US astronaut aboard the International Space Station (ISS) recently caught a glimpse of one of Earth's least understood atmospheric phenomena. While orbiting in the early hours of July 3, Nichole "Vapor" Ayers snapped a photo of a transient luminous event, as she passed over North America. Better known as a sprite, these atmospheric events are common after a lightning strike. Wow," Ayers posted to social media later that day along with the stunning picture.

Homograph disambiguation, the task of distinguishing words with identical spellings but different meanings, poses a substantial challenge in natural language processing. In this study, we introduce a novel dataset tailored for Persian homograph disambiguation. Our work encompasses a thorough exploration of various embeddings, evaluated through the cosine similarity method and their efficacy in downstream tasks like classification. Our investigation entails training a diverse array of lightweight machine learning and deep learning models for phonograph disambiguation. We scrutinize the models' performance in terms of Accuracy, Recall, and F1 Score, thereby gaining insights into their respective strengths and limitations. The outcomes of our research underscore three key contributions. First, we present a newly curated Persian dataset, providing a solid foundation for future research in homograph disambiguation. Second, our comparative analysis of embeddings highlights their utility in different contexts, enriching the understanding of their capabilities. Third, by training and evaluating a spectrum of models, we extend valuable guidance for practitioners in selecting suitable strategies for homograph disambiguation tasks. In summary, our study unveils a new dataset, scrutinizes embeddings through diverse perspectives, and benchmarks various models for homograph disambiguation. These findings empower researchers and practitioners to navigate the intricate landscape of homograph-related challenges effectively.

Generative AI's ability to harness customer data in a highly sophisticated manner means enterprises are accelerating plans to invest in and leverage the technology's capabilities. In a study titled "The Future of Enterprise Data & AI," Corinium Intelligence and WNS Triange surveyed 100 global C-suite leaders and decision-makers specializing in AI, analytics, and data. Seventy-six percent of the respondents said that their organizations are already using or planning to use generative AI. According to McKinsey, while generative AI will affect most business functions, "four of them will likely account for 75% of the total annual value it can deliver." Among these are marketing and sales and customer operations.

As a fourth-year ophthalmology resident at Emory University School of Medicine, Dr. Riley Lyons' biggest responsibilities include triage: When a patient comes in with an eye-related complaint, Lyons must make an immediate assessment of its urgency. He often finds patients have already turned to "Dr. Online, Lyons said, they are likely to find that "any number of terrible things could be going on based on the symptoms that they're experiencing." So, when two of Lyons' fellow ophthalmologists at Emory came to him and suggested evaluating the accuracy of the AI chatbot ChatGPT in diagnosing eye-related complaints, he jumped at the chance. In June, Lyons and his colleagues reported in medRxiv, an online publisher of preliminary health science studies, that ChatGPT compared quite well to human doctors who reviewed the same symptoms -- and performed vastly better than the symptom checker on the popular health website WebMD. And despite the much-publicized "hallucination" problem known to ...

Chris Winfield, founder of Understanding A.I., tells'Fox & Friends Weekend' host Will Cain about a study showing patients preferred medical answers from artificial intelligence over doctors. When it comes to answering medical questions, can ChatGPT do a better job than human doctors? It appears to be possible, according to the results of a new study published in JAMA Internal Medicine, led by researchers from the University of California San Diego. The researchers compiled a random sample of nearly 200 medical questions that patients posted on Reddit, a popular social discussion website, for doctors to answer. Next, they entered the questions into ChatGPT (OpenAI's artificial intelligence chatbot) and recorded its response.

A Chicago native was on a mission to prove the $30,000 painting he purchased at an antique store was created by the famed Raphael and 30 years later, artificial intelligence finds a 97 percent match. Anthony Ayers bought the painting, known as the Flaget Madonna, in 1995 and four years ago, he commissioned an AI firm to analyze the paint and wood panels. Art Recognition used its machine learning software to examine the brushstrokes, determining the famed Italian painted the faces of Mary and infant Jesus with high probability. The firm trained its software with millions of artworks to verify the authenticity of art, and fewer than 10 percent of its clients have received higher than 95 percent. While Ayers and his friends have spent more than $500,000 hiring specialists to analyze the paints and wooden panels, the last Raphael masterpiece sold for $48 million in 2009 - a sketch titled'Head of a Muse.' Ayers, however, died in 2022 at age 64, but his wife, Dawn Turco, is carrying on with her late husband's quest.

In March, when Netflix quietly dropped its original teen suicide mystery series 13 Reasons Why, it took a few days for people to start freaking out. But soon, schools started sending home notes warning parents about the show's graphic depictions of suicide and rape. Psychologists wrote op-eds denouncing its disregard for the World Health Organization's suicide portrayal guidelines. News outlets published more than 600,000 stories about it. And then, there was Twitter.