Krishna C Puvvada, Faisal Ladhak, Santiago Akle Serano, Cheng-Ping Hsieh, Shantanu Acharya, Somshubra Majumdar, Fei Jia, Samuel Kriman, Simeng Sun, Dima Rekesh, Boris Ginsburg. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Inference with Transformer-based Large Language Models (LLMs) on long sequences is both costly and slow due to the quadratic complexity of the self-attention mechanism. We introduce Star Attention, a two-phase block-sparse approximation that improves computational efficiency by sharding attention across multiple hosts while minimizing communication overhead. In the first phase, the context is processed using blockwise-local attention across hosts, in parallel. In the second phase, query and response tokens attend to all prior cached tokens through sequence-global attention. Star Attention integrates seamlessly with most Transformer-based LLMs trained with global attention, reducing memory requirements and inference time by up to 11x while preserving 97-100% of accuracy.

—Video surveillance via closed-circuit television (CCTV) is ubiquitous across the world today, with a multiplicity of monitoring objectives. However, IoT-enabled real-time automated detection of fire and smoke has not been a part of the CCTV repertoire so far. The present study addresses the feasibility and establishes a methodology for detecting fire and smoke solely from live video streams using a novel IoT-centric framework with minimal latency and higher accuracy, and without the need for additional data, such as from thermal or multi-spectral imaging streams. In our design, CCTV/IP cameras serve purely as Perception Layer nodes that stream raw video (without ondevice inference or pre-processing), transmitting via IoT network protocols (RTSP/RTMP/SDP) to a centralized Application Layer for AI inference. It is shown that available artificial intelligence (AI) object detection models (EfficientDet, Faster-RCNN, YOLO, and their variants) that have been trained and tested with a comprehensive dataset of fire and smoke images are able to detect fire and smoke from video streams with varied accuracies that make the stand-alone models unreliable for practical applications. By uniquely integrating multiple AI algorithms into an end-toend IoT architecture (Perception–Network–Application layers), as well as incorporating a novel technique of monitoring and analyzing the area of dynamically assigned bounding boxes around detected fire and smoke regions, the overall accuracy is significantly increased, creating a robust system for detecting all classes of fire within the same system. The dynamic analysis of bounding box areas over consecutive frames allows the system to differentiate between real, live fires, and static images or objects that resemble fire, thereby significantly reducing false positives. The framework leverages IoT camera nodes streaming via RTSP/RTMP/SDP to cloud-hosted ingestion servers on Amazon Web Service (AWS) EC2 (Application Layer), where video frames are continuously extracted and fed to an AI ensemble. Detected events, validated by multimodel consensus, are stored in S3 and trigger alerts in real-time via AWS SNS/SES based on custom risk thresholds. The [email protected] and detection time of Scaled-YOLOv4 (80.6%, 0.016 seconds/video frame) and EfficientDet (78.1%, 0.019 seconds/video frame) are found to be better than that of Faster-RCNN (67.8%, 0.054 seconds/video frame), highlighting the applicability of Scaled-YOLOv4 and EfficientDet for highthroughput, low-latency real-time detections within IoT-driven surveillance ecosystems.

Automatic speech recognition models are often adapted to improve their accuracy in a new domain. A potential drawback of model adaptation to new domains is catastrophic forgetting, where the Word Error Rate on the original domain is significantly degraded. This paper addresses the situation when we want to simultaneously adapt automatic speech recognition models to a new domain and limit the degradation of accuracy on the original domain without access to the original training dataset. We propose several techniques such as a limited training strategy and regularized adapter modules for the Transducer encoder, prediction, and joiner network. We apply these methods to the Google Speech Commands and to the UK and Ireland English Dialect speech data set and obtain strong results on the new target domain while limiting the degradation on the original domain.

We introduce the use of Poincaré embeddings to improve existing state-of-the-art approaches to domain-specific taxonomy induction from text as a signal for both relocating wrong hyponym terms within a (pre-induced) taxonomy as well as for attaching disconnected terms in a taxonomy.This method substantially improves previous state-of-the-art results on the SemEval-2016 Task 13 on taxonomy extraction.We demonstrate the superiority of Poincaré embeddings over distributional semantic representations, supporting the hypothesis that they can better capture hierarchical lexical-semantic relationships than embeddings in the Euclidean space.