Yang Jin

Yang Jin, Zhicheng Sun, Kun Xu et al. · pku

In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.

Peijie Qiu, Jin Yang, Sayantan Kumar et al.

In the past decades, deep neural networks, particularly convolutional neural networks, have achieved state-of-the-art performance in a variety of medical image segmentation tasks. Recently, the introduction of the vision transformer (ViT) has significantly altered the landscape of deep segmentation models. There has been a growing focus on ViTs, driven by their excellent performance and scalability. However, we argue that the current design of the vision transformer-based UNet (ViT-UNet) segmentation models may not effectively handle the heterogeneous appearance (e.g., varying shapes and sizes) of objects of interest in medical image segmentation tasks. To tackle this challenge, we present a structured approach to introduce spatially dynamic components to the ViT-UNet. This adaptation enables the model to effectively capture features of target objects with diverse appearances. This is achieved by three main components: \textbf{(i)} deformable patch embedding; \textbf{(ii)} spatially dynamic multi-head attention; \textbf{(iii)} deformable positional encoding. These components were integrated into a novel architecture, termed AgileFormer. AgileFormer is a spatially agile ViT-UNet designed for medical image segmentation. Experiments in three segmentation tasks using publicly available datasets demonstrated the effectiveness of the proposed method. The code is available at \href{https://github.com/sotiraslab/AgileFormer}{https://github.com/sotiraslab/AgileFormer}.

Jin Yang, Daniel S. Marcus, Aristeidis Sotiras

U-Net has been widely used for segmenting abdominal organs, achieving promising performance. However, when it is used for multi-organ segmentation, first, it may be limited in exploiting global long-range contextual information due to the implementation of standard convolutions. Second, the use of spatial-wise downsampling (e.g., max pooling or strided convolutions) in the encoding path may lead to the loss of deformable or discriminative details. Third, features upsampled from the higher level are concatenated with those that persevered via skip connections. However, repeated downsampling and upsampling operations lead to misalignments between them and their concatenation degrades segmentation performance. To address these limitations, we propose Dynamically Calibrated Convolution (DCC), Dynamically Calibrated Downsampling (DCD), and Dynamically Calibrated Upsampling (DCU) modules, respectively. The DCC module can utilize global inter-dependencies between spatial and channel features to calibrate these features adaptively. The DCD module enables networks to adaptively preserve deformable or discriminative features during downsampling. The DCU module can dynamically align and calibrate upsampled features to eliminate misalignments before concatenations. We integrated the proposed modules into a standard U-Net, resulting in a new architecture, termed Dynamic U-Net. This architectural design enables U-Net to dynamically adjust features for different organs. We evaluated Dynamic U-Net in two abdominal multi-organ segmentation benchmarks. Dynamic U-Net achieved statistically improved segmentation accuracy compared with standard U-Net. Our code is available at https://github.com/sotiraslab/DynamicUNet.

Peijie Qiu, Wenhui Zhu, Sayantan Kumar et al.

Multiple signal modalities, such as vision and sounds, are naturally present in real-world phenomena. Recently, there has been growing interest in learning generative models, in particular variational autoencoder (VAE), to for multimodal representation learning especially in the case of missing modalities. The primary goal of these models is to learn a modality-invariant and modality-specific representation that characterizes information across multiple modalities. Previous attempts at multimodal VAEs approach this mainly through the lens of experts, aggregating unimodal inference distributions with a product of experts (PoE), a mixture of experts (MoE), or a combination of both. In this paper, we provide an alternative generic and theoretical formulation of multimodal VAE through the lens of barycenter. We first show that PoE and MoE are specific instances of barycenters, derived by minimizing the asymmetric weighted KL divergence to unimodal inference distributions. Our novel formulation extends these two barycenters to a more flexible choice by considering different types of divergences. In particular, we explore the Wasserstein barycenter defined by the 2-Wasserstein distance, which better preserves the geometry of unimodal distributions by capturing both modality-specific and modality-invariant representations compared to KL divergence. Empirical studies on three multimodal benchmarks demonstrated the effectiveness of the proposed method.

Jin Yang, Qiong Wu, Zhiying Feng et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities, leading to a significant increase in user demand for LLM services. However, cloud-based LLM services often suffer from high latency, unstable responsiveness, and privacy concerns. Therefore, multiple LLMs are usually deployed at the network edge to boost real-time responsiveness and protect data privacy, particularly for many emerging smart mobile and IoT applications. Given the varying response quality and latency of LLM services, a critical issue is how to route user requests from mobile and IoT devices to an appropriate LLM service (i.e., edge LLM expert) to ensure acceptable quality-of-service (QoS). Existing routing algorithms fail to simultaneously address the heterogeneity of LLM services, the interference among requests, and the dynamic workloads necessary for maintaining long-term stable QoS. To meet these challenges, in this paper we propose a novel deep reinforcement learning (DRL)-based QoS-aware LLM routing framework for sustained high-quality LLM services. Due to the dynamic nature of the global state, we propose a dynamic state abstraction technique to compactly represent global state features with a heterogeneous graph attention network (HAN). Additionally, we introduce an action impact estimator and a tailored reward function to guide the DRL agent in maximizing QoS and preventing latency violations. Extensive experiments on both Poisson and real-world workloads demonstrate that our proposed algorithm significantly improves average QoS and computing resource efficiency compared to existing baselines.