GPT detectors frequently misclassify non-native English writing as AI generated, raising concerns about fairness and robustness. Addressing the biases in these is crucial to prevent marginalization of speakers evaluative educational settings create a more equitable digital landscape.

There are now over 500 medical artificial intelligence (AI) devices that approved by the U.S. Food and Drug Administration. However, little is known about where how often these actually used after regulatory approval. In this article, we systematically quantify adoption usage of AI in United States tracking Current Procedural Terminology (CPT) codes explicitly created for AI. CPT widely documenting billing payment procedures, providing a measure device utilization across different clinical...

We present an approach for estimating the fraction of text in a large corpus which is likely to be substantially modified or produced by language model (LLM). Our maximum likelihood leverages expert-written and AI-generated reference texts accurately efficiently examine real-world LLM-use at level. apply this case study scientific peer review AI conferences that took place after release ChatGPT: ICLR 2024, NeurIPS 2023, CoRL 2023 EMNLP 2023. results suggest between 6.5% 16.9% submitted as...

Scientific publishing lays the foundation of science by disseminating research findings, fostering collaboration, encouraging reproducibility, and ensuring that scientific knowledge is accessible, verifiable, built upon over time. Recently, there has been immense speculation about how many people are using large language models (LLMs) like ChatGPT in their academic writing, to what extent this tool might have an effect on global practices. However, we lack a precise measure proportion...

Expert feedback lays the foundation of rigorous research. However, rapid growth scholarly production and intricate knowledge specialization challenge conventional scientific mechanisms. High-quality peer reviews are increasingly difficult to obtain. Researchers who more junior or from under-resourced settings have especially hard times getting timely feedback. With breakthrough large language models (LLM) such as GPT-4, there is growing interest in using LLMs generate on research...

A major bottleneck in training end-to-end task-oriented dialog system is the lack of data. To utilize limited data more efficiently, we propose Modular Supervision Network (MOSS), an encoder-decoder framework that could incorporate supervision from various intermediate modules including natural language understanding, state tracking, policy learning and generation. With only 60% data, MOSS-all (i.e., MOSS with all four modules) outperforms state-of-the-art models on CamRest676. Moreover,...

Recent advancements in deep learning techniques facilitate intelligent-query support diverse applications, such as content-based image retrieval and audio texturing. Unlike conventional key-based queries, these intelligent queries lack efficient indexing require complex compute operations for feature matching. To achieve high-performance querying against massive datasets, modern computing systems employ GPUs in-conjunction with solid-state drives (SSDs) fast data access parallel processing....

Training a supervised neural network classifier typically requires many annotated training samples. Collecting and annotating large number of data points are costly sometimes even infeasible. Traditional annotation process uses low-bandwidth human-machine communication interface: classification labels, each which only provides few bits information. We propose Active Learning with Contrastive Explanations (ALICE), an expert-in-the-loop framework that utilizes contrastive natural language...

Images are more than a collection of objects or attributes — they represent web relationships among interconnected objects. Scene Graph has emerged as new modality structured graphical representation images. encodes nodes connected via pairwise relations edges. To support question answering on scene graphs, we propose GraphVQA, language-guided graph neural network framework that translates and executes natural language multiple iterations message passing nodes. We explore the design space...

Open Domain dialog system evaluation is one of the most important challenges in research. Existing automatic metrics, such as BLEU are mostly reference-based. They calculate difference between generated response and a limited number available references. Likert-score based self-reported user rating widely adopted by social conversational systems, Amazon Alexa Prize chatbots. However, suffers from bias variance among different users. To alleviate this problem, we formulate comparison task. We...

State Space Models (SSMs) have emerged as efficient alternatives to Transformers for sequential modeling, but their inability leverage modality-specific features limits performance in multi-modal pretraining. Here, we propose Mixture-of-Mamba, a novel SSM architecture that introduces modality-aware sparsity through parameterization of the Mamba block. Building on Mixture-of-Transformers (W. Liang et al. arXiv:2411.04996; 2024), extend benefits SSMs while preserving computational efficiency....

ML libraries, often written in architecture-specific programming languages (ASPLs) that target domain-specific architectures, are key to efficient systems. However, writing these high-performance libraries is challenging because it requires expert knowledge of algorithms and the ASPL. Large language models (LLMs), on other hand, have shown general coding capabilities. challenges remain when using LLMs for generating ASPLs 1) this task complicated even experienced human programmers 2) there...

Training a Generative Adversarial Networks (GAN) for new domain from scratch requires an enormous amount of training data and days time. To this end, we propose DAWSON, Domain Adaptive FewShot Generation FrameworkFor GANs based on meta-learning. A major challenge applying meta-learning is to obtain gradients the generator evaluating it development sets due likelihood-free nature GANs. address challenge, alternative GAN procedure that naturally combines two-step algorithms. DAWSON...

Recent advances in deep learning have made the use of large, neural networks with tens millions parameters. The sheer size these imposes a challenging computational burden during inference. Existing work focuses primarily on accelerating each forward pass network. Inspired by group testing strategy for efficient disease testing, we propose which accelerates samples one pass. Groups that test negative are ruled out. If tests positive, then retested adaptively. A key challenge is to modify...

This article systematically investigates the technology licensing by Stanford University. We analyzed all inventions marketed Stanford's Office of Technology Licensing (OTL) between 1970 to 2020, with 4,512 from 6,557 inventors. quantified how innovation landscape at changed over time and examined factors that correlate commercial success. found most profitable are predominantly licensed inventors' own startups, have involved larger teams time, proportion female inventors has tripled past 25...

Advances in machine learning are closely tied to the creation of datasets. While data documentation is widely recognized as essential reliability, reproducibility, and transparency ML, we lack a systematic empirical understanding current dataset practices. To shed light on this question, here take Hugging Face -- one largest platforms for sharing collaborating ML models datasets prominent case study. By analyzing all 7,433 Face, our investigation provides an overview ecosystem insights into...

Weixin Liang, Kai-Hui Zhou Yu. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

The predominant approach to visual question answering (VQA) relies on encoding the image and with a "black-box" neural encoder decoding single token as answer like "yes" or "no". Despite this approach's strong quantitative results, it struggles come up intuitive, human-readable forms of justification for prediction process. To address insufficiency, we reformulate VQA full generation task, which requires model justify its predictions in natural language. We propose LRTA [Look, Read, Think,...

The rapid proliferation of AI models has underscored the importance thorough documentation, as it enables users to understand, trust, and effectively utilize these in various applications. Although developers are encouraged produce model cards, it's not clear how much information or what cards contain. In this study, we conduct a comprehensive analysis 32,111 documentations on Hugging Face, leading platform for distributing deploying models. Our investigation sheds light prevailing card...