Edge deployment of large Vision-Language Models (VLMs) increasingly relies on flash-based weight offloading, where activation sparsification is used to reduce I/O overhead. However, conventional sparsification remains model-centric, selecting neurons solely by activation magnitude and neglecting how access patterns influence flash performance. We present Neuron Chunking, an I/O-efficient sparsification strategy that operates on *chunks*—groups of contiguous neurons in memory—and couples neuron importance with storage access cost. The method models I/O latency through a lightweight abstraction of access contiguity and selects chunks with high utility, defined as neuron importance normalized by estimated latency. By aligning sparsification decisions with the underlying storage behavior, Neuron Chunking improves I/O efficiency by up to 4.65× and 5.76× on Jetson Orin Nano and Jetson AGX Orin, respectively. The code is available at https://github.com/snuhcs/vlm-flash.

Cloud-based live video analytics with tight latency bound is gaining importance to support emerging applications such as UAVs and augmented reality. However, existing systems often struggle to meet stringent latency constraints under fluctuating network conditions with packet losses and late-arriving packets. We propose a loss-tolerant live video analytics system called Logan, which effectively accepts packet losses while maintaining high accuracy by utilizing the inherent resilience in DNNs. We design i) Codec-aware Inpainting, which accurately recovers the frame error from packet losses ii) Fast-Forward Recovery that prevents the remaining un-recovered error from propagating over future frames indefinitely. Our results show a 3× improvement (33.2%→99.9%) in SLO satisfaction rate compared to the reliable transmission scheme with <1% accuracy drop under a 5% packet loss rate.

We present FlexPatch, a novel mobile system to enable accurate and real-time object detection over high-resolution video streams. A widely-used approach for real-time video analysis is detection-based tracking (DBT), i.e., running the heavy-but-accurate detector every few frames and applying a lightweight tracker for in-between frames. However, the approach is limited for real-time processing of high-resolution videos in that i) a lightweight tracker fails to handle occlusion, object appearance changes, and occurrences of new objects, and ii) the detection results do not effectively offset tracking errors due to the high detection latency. We propose tracking-aware patching technique to address such limitations of the DBT frameworks. It effectively identifies a set of subareas where the tracker likely fails and tightly packs them into a small-sized rectangular area where the detection can be efficiently performed at low latency. This prevents the accumulation of tracking errors and offsets the tracking errors with frequent fresh detection results. Our extensive evaluation shows that FlexPatch not only enables real-time and power-efficient analysis of high-resolution frames on mobile devices but also improves the overall accuracy by 146% compared to baseline DBT frameworks.

The complete sharing of parameters for multilingual translation (1-1) has been the mainstream approach in current research.However, degraded performance due to the capacity bottleneck and low maintainability hinders its extensive adoption in industries.In this study, we revisit the multilingual neural machine translation model that only share modules among the same languages (M2) as a practical alternative to 1-1 to satisfy industrial requirements.Through comprehensive experiments, we identify the benefits of multi-way training and demonstrate that the M2 can enjoy these benefits without suffering from the capacity bottleneck.Furthermore, the interlingual space of the M2 allows convenient modification of the model.By leveraging trained modules, we find that incrementally added modules exhibit better performance than singly trained models.The zero-shot performance of the added modules is even comparable to supervised models.Our findings suggest that the M2 can be a competent candidate for multilingual translation in industries.