Artificial Intelligence Generated Content (AIGC) models typically have hundreds of billions of parameters, and developers experience prohibitively long pulling time from a central model registry to their local environments. Peer-to-peer (P2P)-enabled model distribution that pulls models from local peers within a cluster instead of the central model registry is emerging as a promising technique to reduce model pulling time. Nonetheless, such model distribution systems have to handle bursty concurrent pulling tasks. This may occupy the network bandwidth of some peers for a long time, thereby making the peers unable to respond to further pulling tasks. We thus aim to design a lightweight and scalable model distribution system to balance the network resource usage of peers, by proposing learning-driven algorithms to accurately predict network status between peers and implementing the design in real production environments. Specifically, we first propose a lightweight network measurement mechanism that combines active delay probing and passive bandwidth inference with low resource overhead. We also propose a learning-driven task scheduling algorithm based on a structural graph representation with a varied-multi-hop attention mechanism, to predict bursty patterns of concurrent pulling tasks. We then design an asynchronous model training and inference method to enable seamless incremental learning based on the dynamic network status data. We finally implement our system design and the learning-driven algorithm in a Cloud Native Computing Foundation (CNCF) projectDragonflythat has already been publicly released since its version$v2.1.0$. Real experiments in the Ant Group’s production environment show that our system reduces the total completion time by at least 10% and increases the average bandwidth utilization of peers by 20%, compared with mainstream systems and algorithms.

Class incremental medical segmentation (CIMS) aims to sequentially learn new classes while preserving knowledge of previously learned categories in the absence of old-class labels. Current methods suffer from performance degradation under class imbalance and require additional segmentation heads to accommodate new categories. Inspired by recent prototype learning that leverages prototypes to achieve robust recognition of new categories under limited-data regimes, we introduce a Prototype-dRIven class increMEntal (PRIME) method. PRIME replaces the incremental segmentation heads with prototypes to mitigate class imbalance, allowing new class learning with the simple addition of new prototypes. Based on prototype learning, PRIME further involves three tailored techniques. First, prototype structure alignment imposes structural constraints on inter-prototype relations to maintain consistent relative distances in the feature space, improving the model's ability to distinguish distinct classes. Second, pixel-wise contrastive loss term groups embeddings of similar samples while separating those of different classes, enhancing segmentation accuracy across all categories. Finally, the consensus-based prototype update mechanism refines the old prototypes during the learning of new classes, preventing performance degradation on the old classes. Extensive experiments on two public multi-organ segmentation datasets demonstrate that our approach significantly outperforms state-of-the-art methods, validating the effectiveness of the proposed PRIME.

Pretrained Foundation Models (PFMs) are regarded as a promising accelerator for the development of various Artificial Intelligence (AI) applications, and have recently been widely fine-tuned to satisfy users' personalized inference demands. As many users are attracted to PFM-based AI applications, remote data centers are increasingly unable to solely bear the enormous computational demands and meet the delay requirements of inference requests. Mobile edge computing (MEC) offers a viable solution for delivering low-latency inference services by pulling fine-tuned PFMs from the remote data center to cloudlets in the proximity of users. However, a fine-tuned PFM typically comprises billions of model parameters, which are highly resource-intensive, time-consuming, and cost-prohibitive to execute at the edge. To address this, we investigate a novel joint large model selection and pulling problem in MEC networks. The novelty of our study lies in exploring parameter sharing among fine-tuned PFMs based on their common knowledge. Specifically, we first formulate a Non-Linear Integer Programming (NLIP) for the problem to minimize the total delay of implementing all inference requests. We then transform the NLIP into an equivalent Integer Linear Program (ILP) that is much simpler to solve. We further propose a randomized algorithm with a provable approximation ratio for the problem. We also consider the online version of the problem with uncertain request demand, and develop an online learning algorithm with a bounded regret. The crux of the online algorithm is the adoption of the multi-armed bandit technique with restricted context for dynamic admissions of inference requests. We finally conduct extensive experiments based on real datasets. Experimental results demonstrate that our algorithms reduce at least 38% in total delays and average costs, while achieving a 5% improvement in average accuracies.

Recent advances in medical image segmentation models combine convolution with the attention mechanism which provides an effective approach to formulate long-term dependencies. However, many works either replaced the convolutional layers with attention layers or embedded attention layers into convolutional neural network (CNN)-based models. To explore the potential of hybrid architecture, we propose a simple cascade module that builds up multiple receptive fields using convolutional kernels with different sizes and learns global context via self-attention layers. Benefiting from the powerful representation ability of the proposed module, multilayer perceptrons (MLPs) with shift operation are adopted to bridge the encoder and decoder to reduce the model size without losing accuracy. Experiments show that our model consistently outperforms the latest 2D and 3D models by large margins on three public tasks and is more resilient to shape, size, and boundary variations. The code is available at https://github.com/cicailalala/AERFNet.

Breast tumor segmentation of ultrasound images provides valuable information of tumors for early detection and diagnosis. Accurate segmentation is challenging due to low image contrast between areas of interest; speckle noises, and large inter-subject variations in tumor shape and size. This paper proposes a novel Multi-scale Dynamic Fusion Network (MDF-Net) for breast ultrasound tumor segmentation. It employs a two-stage end-to-end architecture with a trunk sub-network for multiscale feature selection and a structurally optimized refinement sub-network for mitigating impairments such as noise and inter-subject variation via better feature exploration and fusion. The trunk network is extended from UNet++ with a simplified skip pathway structure to connect the features between adjacent scales. Moreover, deep supervision at all scales, instead of at the finest scale in UNet++, is proposed to extract more discriminative features and mitigate errors from speckle noise via a hybrid loss function. Unlike previous works, the first stage is linked to a loss function of the second stage so that both the preliminary segmentations and refinement subnetworks can be refined together at training. The refinement sub-network utilizes a structurally optimized MDF mechanism to integrate preliminary segmentation information (capturing general tumor shape and size) at coarse scales and explores inter-subject variation information at finer scales. Experimental results from two public datasets show that the proposed method achieves better Dice and other scores over state-of-the-art methods. Qualitative analysis also indicates that our proposed network is more robust to tumor size/shapes, speckle noise and heavy posterior shadows along tumor boundaries. An optional post-processing step is also proposed to facilitate users in mitigating segmentation artifacts. The efficiency of the proposed network is also illustrated on the "Electron Microscopy neural structures segmentation dataset". It outperforms a state-of-the-art algorithm based on UNet-2022 with simpler settings. This indicates the advantages of our MDF-Nets in other challenging image segmentation tasks with small to medium data sizes.

Speech emotion recognition (SER) aims to automatically recognize emotional category for a given speech utterance. The performance of a SER system heavily relies on the effectiveness of global representation expressed at utterance level. To effectively extract such a global feature, the mainstream of recent SER architectures adopts a pipeline with two key modules, feature extraction and aggregation. Although variant module designs have brought impressive progresses, SER is still a challenging task. In contrast with those previous works, herein we propose a novel strategy for global SER feature extraction by applying an additional enhancement module on top of the current SER pipeline. To verify its effect, an end-to-end SER architecture is proposed where stacked multiple transformer layers are explored to enhance the aggregated global feature. Such an architecture is evaluated on IEMO-CAP and results strongly substantiate the effectiveness of our proposal. In terms of weighted accuracy on four emotion categories, our proposed SER system outperforms the prior arts by a large margin of relatively 20% improvement. Our codes and the pre-trained SER models are made publicly available.