Hui Sun

Ziyan Huang, Haoyu Wang, Zhongying Deng et al.

Large-scale models pre-trained on large-scale datasets have profoundly advanced the development of deep learning. However, the state-of-the-art models for medical image segmentation are still small-scale, with their parameters only in the tens of millions. Further scaling them up to higher orders of magnitude is rarely explored. An overarching goal of exploring large-scale models is to train them on large-scale medical segmentation datasets for better transfer capacities. In this work, we design a series of Scalable and Transferable U-Net (STU-Net) models, with parameter sizes ranging from 14 million to 1.4 billion. Notably, the 1.4B STU-Net is the largest medical image segmentation model to date. Our STU-Net is based on nnU-Net framework due to its popularity and impressive performance. We first refine the default convolutional blocks in nnU-Net to make them scalable. Then, we empirically evaluate different scaling combinations of network depth and width, discovering that it is optimal to scale model depth and width together. We train our scalable STU-Net models on a large-scale TotalSegmentator dataset and find that increasing model size brings a stronger performance gain. This observation reveals that a large model is promising in medical image segmentation. Furthermore, we evaluate the transferability of our model on 14 downstream datasets for direct inference and 3 datasets for further fine-tuning, covering various modalities and segmentation targets. We observe good performance of our pre-trained model in both direct inference and fine-tuning. The code and pre-trained models are available at https://github.com/Ziyan-Huang/STU-Net.

Yizhou Chen, Guangda Huzhang, Anxiang Zeng et al.

In recent years, recommender systems have advanced rapidly, where embedding learning for users and items plays a critical role. A standard method learns a unique embedding vector for each user and item. However, such a method has two important limitations in real-world applications: 1) it is hard to learn embeddings that generalize well for users and items with rare interactions on their own; and 2) it may incur unbearably high memory costs when the number of users and items scales up. Existing approaches either can only address one of the limitations or have flawed overall performances. In this paper, we propose Clustered Embedding Learning (CEL) as an integrated solution to these two problems. CEL is a plug-and-play embedding learning framework that can be combined with any differentiable feature interaction model. It is capable of achieving improved performance, especially for cold users and items, with reduced memory cost. CEL enables automatic and dynamic clustering of users and items in a top-down fashion, where clustered entities jointly learn a shared embedding. The accelerated version of CEL has an optimal time complexity, which supports efficient online updates. Theoretically, we prove the identifiability and the existence of a unique optimal number of clusters for CEL in the context of nonnegative matrix factorization. Empirically, we validate the effectiveness of CEL on three public datasets and one business dataset, showing its consistently superior performance against current state-of-the-art methods. In particular, when incorporating CEL into the business model, it brings an improvement of $+0.6\%$ in AUC, which translates into a significant revenue gain; meanwhile, the size of the embedding table gets $2650$ times smaller.

Richard Archibald, Feng Bao, Yanzhao Cao et al.

In this paper, we carry out numerical analysis to prove convergence of a novel sample-wise back-propagation method for training a class of stochastic neural networks (SNNs). The structure of the SNN is formulated as discretization of a stochastic differential equation (SDE). A stochastic optimal control framework is introduced to model the training procedure, and a sample-wise approximation scheme for the adjoint backward SDE is applied to improve the efficiency of the stochastic optimal control solver, which is equivalent to the back-propagation for training the SNN. The convergence analysis is derived with and without convexity assumption for optimization of the SNN parameters. Especially, our analysis indicates that the number of SNN training steps should be proportional to the square of the number of layers in the convex optimization case. Numerical experiments are carried out to validate the analysis results, and the performance of the sample-wise back-propagation method for training SNNs is examined by benchmark machine learning examples.

Jinming Xing, Guoheng Sun, Hui Sun et al.

Spatial-temporal network traffic forecasting is a challenging task due to the complex spatial relationships and dynamic temporal patterns present in each node. Traditional regression methods are not directly applicable to such graph data. Recently, Graph Neural Networks (GNNs) have been widely used to model spatial-temporal dependencies. However, existing methods face several limitations: (1) They rely solely on a predefined spatial adjacency matrix, overlooking hidden low-level temporal information. (2) They model spatial and temporal information separately, which inevitably leads to a loss of joint dependencies, or they capture only global or local dependencies. To address these issues, we propose the \textbf{G}lobal-\textbf{L}ocal \textbf{S}patial-\textbf{T}emporal \textbf{a}ware \textbf{G}raph \textbf{AT}tention Network (GLSTaGAT). Specifically, we adopt a data-driven spatial-temporal fusion graph that incorporates low-level spatial and temporal information, serving as the foundation for further graph convolutions. The GLSTaGAT block and its pooling variant are proposed to simultaneously capture local and global spatial-temporal dependencies. Additionally, we introduce node normalization to mitigate covariance shifts, enabling a smoother training process. An encoder-only transformer is utilized to model high-level joint dependencies, and a multi-head attention prediction layer is designed for final information aggregation and prediction. Experimental results on real-world datasets demonstrate that GLSTaGAT outperforms the baselines by 32.14\% (MAE), 28.30\% (RMSE), and 20.47\% (SMAPE) on average.