Large-scale pretrained AI models have shown state-of-the-art accuracy in a series of important applications. As the size of pretrained AI models grows dramatically each year in an effort to achieve higher accuracy, training such models requires massive computing and memory capabilities, which accelerates the convergence of AI and HPC. However, there are still gaps in deploying AI applications on HPC systems, which need application and system co-design based on specific hardware features. To this end, this paper proposes BaGuaLu¹, the first work targeting training brain scale models on an entire exascale supercomputer, the New Generation Sunway Supercomputer. By combining hardware-specific intra-node optimization and hybrid parallel strategies, BaGuaLu enables decent performance and scalability on unprecedentedly large models. The evaluation shows that BaGuaLu can train 14.5-trillion-parameter models with a performance of over 1 EFLOPS using mixed-precision and has the capability to train 174-trillion-parameter models, which rivals the number of synapses in a human brain.

Deep neural networks have demonstrated impressive results in medical image analysis, but designing suitable architectures for each specific task is expertise-dependent and time-consuming. Neural architecture search (NAS) offers an effective means of discovering architectures. It has been highly successful in numerous applications, particularly in natural image classification. Yet, medical images possess unique characteristics, such as small regions and a wide variety of lesion sizes, that differentiate them from natural images. Furthermore, most current NAS methods struggle with high computational costs, especially when dealing with high-resolution image datasets. In this paper, we present a novel evolutionary neural architecture search method called Multi-Scale Training-Free Neural Architecture Search to address these challenges. Specifically, to accommodate the broad range of lesion region sizes in disease diagnosis, we develop a new reduction cell search space that enables the search algorithm to explicitly identify the optimal scale combination for multi-scale feature extraction. To overcome the issue of high computational costs, we utilize training-free indicators as performance measures for candidate architectures, which allows us to search for the optimal architecture more efficiently. More specifically, by considering the capability and simplicity of various networks, we formulate a multi-objective optimization problem that involves two training-free indicators and model complexity for candidate architectures. Extensive experiments on a large medical image benchmark and a publicly available breast cancer detection dataset are conducted. The empirical results demonstrate that our MSTF-NAS outperforms both human-designed architectures and current state-of-the-art NAS algorithms on both datasets, indicating the effectiveness of our proposed method.

Automated detection of skin lesions offers excellent potential for interpretative diagnosis and precise treatment of acne vulgar. However, the blurry boundary and small size of lesions make it challenging to detect acne lesions with traditional object detection methods. To better understand the acne detection task, we construct a new benchmark dataset named AcneSCU, consisting of 276 facial images with 31777 instance-level annotations from clinical dermatology. To the best of our knowledge, AcneSCU is the first acne dataset with high-resolution imageries, precise annotations, and fine-grained lesion categories, which enables the comprehensive study of acne detection. More importantly, we propose a novel method called Spatial Aware Region Proposal Network (SA-RPN) to improve the proposal quality of two-stage detection methods. Specifically, the representation learning for the classification and localization task is disentangled with a double head component to promote the proposals for hard samples. Then, Normalized Wasserstein Distance of each proposal is predicted to improve the correlation between the classification scores and the proposals' intersection-over-unions (IoUs). SA-RPN can serve as a plug-and-play module to enhance standard two-stage detectors. Extensive experiments are conducted on both AcneSCU and the public dataset ACNE04, and the results show that the proposed method can consistently outperform state-of-the-art methods. Code and the collected dataset are made available at https://github.com/pingguokiller/acnedetection to stimulate the future research in the acne healthcare community.

Data augmentation is an effective technique to enrich the training data’s diversity and reduce the risk of overfitting. However, different datasets have distinct preferences on various augmentation techniques. Recently, automated data augmentation (auto-augmentation), which could engineer augmentation policy automatically, drew a growing interest. Previous auto-augmentation methods usually utilize a Density Matching strategy, which highly depends on a large data scale to ensure a precise policy evaluation. When facing small-scale datasets, it usually achieves an inferior performance. To address the problem, an improved method named Augmented Density Matching is proposed by augmenting the train data with policies uniformly sampled from a prior distribution, making the policy evaluation more precise. Moreover, we propose a unified search framework for data augmentation and neural architecture (USAA) by formulating the search processes with one formulation. As a result, both optimal augmentation policy and neural architecture could be obtained within one round of search process. Extensive experiments have been conducted on a bag of medical datasets with general small scales, and the results show that the proposed Augmented Density Matching and USAA can outperform the state-of-the-art auto-augmentation and AutoML approaches, respectively.

Change detection (CD) in remote sensing imagery remains challenging due to boundary ambiguity and false alarms caused by high foreground-background similarity and insufficient difference representation. To address these issues, we propose an Edge-Guided Difference Enhancement Network (EGDENet). EGDENet integrates an edge-aware Adaptive Enhancement Module (EAEM) to extract high-frequency edge cues across scales, and a Channel-Spatial Cooperative Difference Module (CSCDM) to refine change features by jointly leveraging spatial and channel-wise differences. An Up-sampling Feature Fusion (UFF) further enhances robustness to scale variations and improves region consistency. Extensive experiments on two public datasets demonstrate that EGDENet achieves superior performance with clearer boundaries compared to state-of-the-art methods. Our source code is publicly available at https://github.com/adleess/-EGDENet.

We propose a novel approach to detect sparse and stable image features by symmetric properties extracted from the visual data. The regional features are formed by a fast qualitative symmetry operator in combination with quantitative symmetry range information. We apply a simple color histogram descriptor to match pre-selected features to those features acquired by our omnidirectional vision system at run time. The complete algorithm produces regional symmetry-based features that are sparse and highly robust to scale change and panoramic image warp, in particular. We present the algorithms of symmetry and feature processing and show their application in an object classification experiment using our platform, the Bremen autonomous wheelchair "Holland III"