Hoang Le

Hoang Le, Liang Zhang, Amir Said et al.

Realizing the potential of neural video codecs on mobile devices is a big technological challenge due to the computational complexity of deep networks and the power-constrained mobile hardware. We demonstrate practical feasibility by leveraging Qualcomm's technology and innovation, bridging the gap from neural network-based codec simulations running on wall-powered workstations, to real-time operation on a mobile device powered by Snapdragon technology. We show the first-ever inter-frame neural video decoder running on a commercial mobile phone, decoding high-definition videos in real-time while maintaining a low bitrate and high visual quality.

Reza Pourreza, Hoang Le, Amir Said et al.

In video compression, coding efficiency is improved by reusing pixels from previously decoded frames via motion and residual compensation. We define two levels of hierarchical redundancy in video frames: 1) first-order: redundancy in pixel space, i.e., similarities in pixel values across neighboring frames, which is effectively captured using motion and residual compensation, 2) second-order: redundancy in motion and residual maps due to smooth motion in natural videos. While most of the existing neural video coding literature addresses first-order redundancy, we tackle the problem of capturing second-order redundancy in neural video codecs via predictors. We introduce generic motion and residual predictors that learn to extrapolate from previously decoded data. These predictors are lightweight, and can be employed with most neural video codecs in order to improve their rate-distortion performance. Moreover, while RGB is the dominant colorspace in neural video coding literature, we introduce general modifications for neural video codecs to embrace the YUV420 colorspace and report YUV420 results. Our experiments show that using our predictors with a well-known neural video codec leads to 38% and 34% bitrate savings in RGB and YUV420 colorspaces measured on the UVG dataset.

Ties van Rozendaal, Tushar Singhal, Hoang Le et al.

Neural video codecs have recently become competitive with standard codecs such as HEVC in the low-delay setting. However, most neural codecs are large floating-point networks that use pixel-dense warping operations for temporal modeling, making them too computationally expensive for deployment on mobile devices. Recent work has demonstrated that running a neural decoder in real time on mobile is feasible, but shows this only for 720p RGB video. This work presents the first neural video codec that decodes 1080p YUV420 video in real time on a mobile device. Our codec relies on two major contributions. First, we design an efficient codec that uses a block-based motion compensation algorithm available on the warping core of the mobile accelerator, and we show how to quantize this model to integer precision. Second, we implement a fast decoder pipeline that concurrently runs neural network components on the neural signal processor, parallel entropy coding on the mobile GPU, and warping on the warping core. Our codec outperforms the previous on-device codec by a large margin with up to 48% BD-rate savings, while reducing the MAC count on the receiver side by $10 \times$. We perform a careful ablation to demonstrate the effect of the introduced motion compensation scheme, and ablate the effect of model quantization.

Amir Said, Reza Pourreza, Hoang Le

Neural-based image and video codecs are significantly more power-efficient when weights and activations are quantized to low-precision integers. While there are general-purpose techniques for reducing quantization effects, large losses can occur when specific entropy coding properties are not considered. This work analyzes how entropy coding is affected by parameter quantizations, and provides a method to minimize losses. It is shown that, by using a certain type of coding parameters to be learned, uniform quantization becomes practically optimal, also simplifying the minimization of code memory requirements. The mathematical properties of the new representation are presented, and its effectiveness is demonstrated by coding experiments, showing that good results can be obtained with precision as low as 4~bits per network output, and practically no loss with 8~bits.

Mehran Taghian, Yunke Peng, Xing Huang et al.

Large foundation models have become central to modern machine learning, with performance scaling predictably with model size and data. However, training and deploying such models incur substantial computational and memory costs, motivating the development of low-precision training techniques. Recent work has demonstrated that 4-bit floating-point (FP4) formats--such as MXFP4 and NVFP4--can be successfully applied to linear GEMM operations in large language models (LLMs), achieving up to 4x improvements in compute throughput and memory efficiency compared to higher-precision baselines. In this work, we investigate the recently proposed HiFloat4 FP4 format for Huawei Ascend NPUs and systematically compare it with MXFP4 in large-scale training settings. All experiments are conducted on Ascend NPU clusters, with linear and expert GEMM operations performed entirely in FP4 precision. We evaluate both dense architectures (e.g., Pangu and LLaMA-style models) and mixture-of-experts (MoE) models, where both standard linear layers and expert-specific GEMMs operate in FP4. Furthermore, we explore stabilization techniques tailored to FP4 training that significantly reduce numerical degradation, maintaining relative error within 1% of full-precision baselines while preserving the efficiency benefits of 4-bit computation. Our results provide a comprehensive empirical study of FP4 training on NPUs and highlight the practical trade-offs between FP4 formats in large-scale dense and MoE models.

Shweta Mahajan, Shreya Kadambi, Hoang Le et al.

We introduce the Do-Undo task and benchmark to address a critical gap in vision-language models: understanding and generating physically plausible scene transformations driven by real-world actions. Unlike prior work focused on object-level edits, Do-Undo requires models to simulate the outcome of a physical action and then accurately reverse it, reflecting true cause-and-effect in the visual world. We curate a large-scale dataset of reversible actions from real-world videos and design a training strategy enforcing consistency for robust action grounding. Our experiments reveal that current models struggle with physical reversibility, underscoring the importance of this task for embodied AI, robotics, and physics-aware generative modeling. Do-Undo establishes an intuitive testbed for evaluating and advancing physical reasoning in multimodal systems.

Abhijith Punnappurath, Luxi Zhao, Hoang Le et al.

RAW images are unprocessed camera sensor output with sensor-specific RGB values based on the sensor's color filter spectral sensitivities. RAW images also incur strong color casts due to the sensor's response to the spectral properties of scene illumination. The sensor- and illumination-specific nature of RAW images makes it challenging to capture RAW datasets for deep learning methods, as scenes need to be captured for each sensor and under a wide range of illumination. Methods for illumination augmentation for a given sensor and the ability to map RAW images between sensors are important for reducing the burden of data capture. To explore this problem, we introduce the first-of-its-kind dataset comprising carefully captured scenes under a wide range of illumination. Specifically, we use a customized lightbox with tunable illumination spectra to capture several scenes with different cameras. Our illumination and sensor mapping dataset has 390 illuminations, four cameras, and 18 scenes. Using this dataset, we introduce a lightweight neural network approach for illumination and sensor mapping that outperforms competing methods. We demonstrate the utility of our approach on the downstream task of training a neural ISP. Link to project page: https://github.com/SamsungLabs/illum-sensor-mapping.

Qiqi Hou, Randall Rauwendaal, Zifeng Li et al.

Recently, 3D Gaussian Splatting (3DGS) has emerged as a significant advancement in 3D scene reconstruction, attracting considerable attention due to its ability to recover high-fidelity details while maintaining low complexity. Despite the promising results achieved by 3DGS, its rendering performance is constrained by its dependence on costly non-commutative alpha-blending operations. These operations mandate complex view dependent sorting operations that introduce computational overhead, especially on the resource-constrained platforms such as mobile phones. In this paper, we propose Weighted Sum Rendering, which approximates alpha blending with weighted sums, thereby removing the need for sorting. This simplifies implementation, delivers superior performance, and eliminates the "popping" artifacts caused by sorting. Experimental results show that optimizing a generalized Gaussian splatting formulation to the new differentiable rendering yields competitive image quality. The method was implemented and tested in a mobile device GPU, achieving on average $1.23\times$ faster rendering.

Duc N. Do, Minh N. Do, Dang Nguyen et al.

Transthoracic echocardiography is the reference standard for confirming structural heart disease (SHD), but first-line screening is limited by cost, workflow burden, and specialist availability. We evaluated whether open pretrained electrocardiogram (ECG) foundation models can support echo-confirmed multi-label SHD detection using the public EchoNext Mini-Model benchmark. Six echocardiography-derived abnormalities were targeted: reduced left ventricular ejection fraction, increased left ventricular wall thickness, aortic stenosis, mitral regurgitation, tricuspid regurgitation, and right ventricular systolic dysfunction. Under a common pipeline, we compared engineered ECG features with gradient boosting, end-to-end waveform learning from scratch, and transfer from open ECG foundation models. We then applied in-domain self-supervised adaptation of an ECG foundation model (ECG-FM) on EchoNext waveforms followed by selective supervised fine-tuning, and evaluated trade-offs between discrimination and adaptation cost. Adapted ECG-FM models achieved the best overall performance: peak macro-AUROC 0.8509 and macro-AUPRC 0.4297, while a parameter-efficient operating point preserved AUROC (0.8501) and attained the highest fixed-threshold macro-F1 0.3691. Late fusion with covariates did not improve threshold-independent discrimination, and evaluated LoRA, alternative backbones, and mixture-of-foundations strategies did not surpass the best adapted single-backbone models. These results indicate that for ECG-based case finding and echocardiography triage, combining target-domain self-supervised adaptation with selective supervised updating of a pretrained ECG backbone is the most effective transfer strategy.

Qiqi Hou, Farzad Farhadzadeh, Amir Said et al.

The rise of new video modalities like virtual reality or autonomous driving has increased the demand for efficient multi-view video compression methods, both in terms of rate-distortion (R-D) performance and in terms of delay and runtime. While most recent stereo video compression approaches have shown promising performance, they compress left and right views sequentially, leading to poor parallelization and runtime performance. This work presents Low-Latency neural codec for Stereo video Streaming (LLSS), a novel parallel stereo video coding method designed for fast and efficient low-latency stereo video streaming. Instead of using a sequential cross-view motion compensation like existing methods, LLSS introduces a bidirectional feature shifting module to directly exploit mutual information among views and encode them effectively with a joint cross-view prior model for entropy coding. Thanks to this design, LLSS processes left and right views in parallel, minimizing latency; all while substantially improving R-D performance compared to both existing neural and conventional codecs.

Shweta Mahajan, Hoang Le, Hyojin Park et al.

Large Vision-Language Models (VLMs) rely on effective multimodal alignment between pre-trained vision encoders and Large Language Models (LLMs) to integrate visual and textual information. This paper presents a comprehensive analysis of attention patterns in efficient VLMs, revealing that concatenation-based architectures frequently fail to distinguish between semantically matching and non-matching image-text pairs. This is a key factor for object hallucination in these models. To address this, we introduce Attention-Guided Efficient Vision-Language Models (AGE-VLM), a novel framework that enhances visual grounding through interleaved cross-attention layers to instill vision capabilities in pretrained small language models. This enforces in VLM the ability "look" at the correct image regions by leveraging spatial knowledge distilled from the Segment Anything Model (SAM), significantly reducing hallucination. We validate our approach across different vision-centric benchmarks where our method is better or comparable to prior work on efficient VLMs. Our findings provide valuable insights for future research aimed at achieving enhanced visual and linguistic understanding in VLMs.

Farzad Farhadzadeh, Qiqi Hou, Hoang Le et al.

Advances in rendering have led to tremendous growth in texture assets, including resolution, complexity, and novel textures components, but this growth in data volume has not been matched by advances in its compression. Meanwhile Neural Image Compression (NIC) has advanced significantly and shown promising results, but the proposed methods cannot be directly adapted to neural texture compression. First, texture compression requires on-demand and real-time decoding with random access during parallel rendering (e.g. block texture decompression on GPUs). Additionally, NIC does not support multi-resolution reconstruction (mip-levels), nor does it have the ability to efficiently jointly compress different sets of texture channels. In this work, we introduce a novel approach to texture set compression that integrates traditional GPU texture representation and NIC techniques, designed to enable random access and support many-channel texture sets. To achieve this goal, we propose an asymmetric auto-encoder framework that employs a convolutional encoder to capture detailed information in a bottleneck-latent space, and at decoder side we utilize a fully connected network, whose inputs are sampled latent features plus positional information, for a given texture coordinate and mip level. This latent data is defined to enable simplified access to multi-resolution data by simply changing the scanning strides. Experimental results demonstrate that this approach provides much better results than conventional texture compression, and significant improvement over the latest method using neural networks.

Florian Tramèr, Reza Shokri, Ayrton San Joaquin et al.

We introduce a new class of attacks on machine learning models. We show that an adversary who can poison a training dataset can cause models trained on this dataset to leak significant private details of training points belonging to other parties. Our active inference attacks connect two independent lines of work targeting the integrity and privacy of machine learning training data. Our attacks are effective across membership inference, attribute inference, and data extraction. For example, our targeted attacks can poison <0.1% of the training dataset to boost the performance of inference attacks by 1 to 2 orders of magnitude. Further, an adversary who controls a significant fraction of the training data (e.g., 50%) can launch untargeted attacks that enable 8x more precise inference on all other users' otherwise-private data points. Our results cast doubts on the relevance of cryptographic privacy guarantees in multiparty computation protocols for machine learning, if parties can arbitrarily select their share of training data.

Hoang Le, Feng Liu, Shu Zhang et al.

Homography estimation is an important step in many computer vision problems. Recently, deep neural network methods have shown to be favorable for this problem when compared to traditional methods. However, these new methods do not consider dynamic content in input images. They train neural networks with only image pairs that can be perfectly aligned using homographies. This paper investigates and discusses how to design and train a deep neural network that handles dynamic scenes. We first collect a large video dataset with dynamic content. We then develop a multi-scale neural network and show that when properly trained using our new dataset, this neural network can already handle dynamic scenes to some extent. To estimate a homography of a dynamic scene in a more principled way, we need to identify the dynamic content. Since dynamic content detection and homography estimation are two tightly coupled tasks, we follow the multi-task learning principles and augment our multi-scale network such that it jointly estimates the dynamics masks and homographies. Our experiments show that our method can robustly estimate homography for challenging scenarios with dynamic scenes, blur artifacts, or lack of textures.

Riccardo Zucca, Xerxes D. Arsiwalla, Hoang Le et al.

We investigate scaling properties of human brain functional networks in the resting-state. Analyzing network degree distributions, we statistically test whether their tails scale as power-law or not. Initial studies, based on least-squares fitting, were shown to be inadequate for precise estimation of power-law distributions. Subsequently, methods based on maximum-likelihood estimators have been proposed and applied to address this question. Nevertheless, no clear consensus has emerged, mainly because results have shown substantial variability depending on the data-set used or its resolution. In this study, we work with high-resolution data (10K nodes) from the Human Connectome Project and take into account network weights. We test for the power-law, exponential, log-normal and generalized Pareto distributions. Our results show that the statistics generally do not support a power-law, but instead these degree distributions tend towards the thin-tail limit of the generalized Pareto model. This may have implications for the number of hubs in human brain functional networks.