Excel to HTML
TUESDAY 3 MARCH 2026
Efficient Programming Imaging Techniques
15:30 - 17:30
Grand Peninsula F
15:30HPCI-203
Differentiable programming for optimization and inverse problems in electron and optical microscopy, [view abstract]
Electron and light optical systems today are both increasingly complex systems, with multiple optical elements and multimodal data streams. This has led to an explosion of data analysis and simulation methods. For modern simulation techniques, the focus has been on parallelization and GPU computing for increased speed and scalability, while machine learning-based approaches are becoming increasingly common for analysis approaches. In this work, we introduce two different packages - ptyrodactyl and janssen for electron and optical microscopy, respectively. The aim of both approaches is to use the autodifferentiation capabilities of JAX for both simulation and inverse problems.
15:50HPCI-204
Code parallelization via large language model, Xiaowen Huang, Shenzhen University (China (Mainland)); Xu Zhang, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (China (Mainland)); Jintao Meng, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (China (Mainland)); Peng Chen, riken; Wenxi Zhu, Tencent (China (Mainland)); Yanjie Wei, Shenzhen Institutes of Advacnecd Technology, Chinese Academy of Sciences (China (Mainland)) [view abstract]
High-performance parallel code generation is a complex and fascinating area in computer science that focuses on producing code that executes as quickly and efficiently as possible. In our paper, we designed a new architecture for parallel code generation agent. It also in-cooperated with two techniques: data augmentation, prompting and retrieval-augmented editing to improve the performance of the parallel codes. Data augmentation is implemented by extracting and processing PIE dataset, and also synthesis dataset generated by LLM models with ParEval benchmark. Finally planning-oriented prompting, code verification and retrieval augmented editing are used to promote the actual performance of the LLM generated code. The evaluation results confirm that a rough speedup of 6.06X and 5.13X are achieved using Qwen2.5-Coder-7B-Instruct, Qwen2.5-Coder-14B-Instruct LLM models.
16:10HPCI-205
Edge-based imaging under power constraints, Xuetao Chen, Hong Kong Baptist University (Hong Kong (Greater China)); Amelie Chi Zhou, Hong Kong Baptist University (Hong Kong (Greater China)); Peng Chen, RIKEN Center for Computational Science (Japan); Du Wu, Institute of Science Tokyo, RIKEN Center for Computational Science (Japan); Edgar Noriega, National Institute of Advanced Industrial Science and Technology (AIST) (Japan); Emmanuel Jeannot, DDN and RIKEN (Japan); Mohamed Wahib, RIKEN Center for Computational Science (Japan) [view abstract]
Computed Tomography (CT) is a vital 3D imaging technology widely used in medicine and research. CT reconstruction on edge devices is challenging due to constraints in compute power, memory, and energy. We propose an optimized framework for Nvidia Jetson devices that leverages both CUDA and Tensor Cores to accelerate the back-projection (BP) kernel with mixed half- and single-precision arithmetic. This approach achieves significantly faster performance on edge devices compared with the baseline and offers higher energy efficiency than high-end GPU servers.
16:30HPCI-206
Lens simulation for high-resolution and multispectral 3D scenes using depth-variant PSF and distributed multi-GPU processing, Seonghyeon Kang, Samsung Electronics Co., Ltd. (Republic of Korea); Jeongyong Shin, Samsung Electronics Co., Ltd. (Republic of Korea); Sangmin Kim, Samsung Electronics Co., Ltd. (Republic of Korea); Jeongwook Lee, Samsung Electronics Co., Ltd. (Republic of Korea); Sung-Su Kim, Samsung Electronics Co., Ltd. (Republic of Korea); Yitae Kim, Samsung Electronics Co., Ltd. (Republic of Korea) [view abstract]
Full optical simulations of 3D scenes using commercial ray-tracing tools such as Speos provide physically accurate results but suffer from excessive computation time. To overcome this limitation, we propose a hybrid simulation pipeline that separates scene rendering and lens modeling into two stages. First, Zemax is used to precompute lens characteristics including distortion maps, depth-variant relative illumination, and a library of depth-dependent point spread functions (PSFs). Second, Speos is employed to render the ideal camera scene and per-pixel depth maps without optical degradation. Finally, a distributed multi-GPU system efficiently applies the Zemax-derived optical degradations to the Speos-rendered images. This approach significantly reduces simulation time while retaining the essential physical properties of lens degradation. The proposed framework provides realistic image formation data for testing computer vision and imaging system design.
Efficient Vision AI Model I
08:30 - 10:30
Grand Peninsula F
08:30HPCI-195
Med-MoSAM: Mixture-of-specialists for efficient and adaptive medical segmentation, Gaofei Chen, University of Alabama at Birmingham (US); Xi Xiao, (US); Hanzhang Chi, University of Alabama at Birmingham (US); Tianyang Wang, University of Alabama at Birmingham (US); Xiao Wang, Oak Ridge National Laboratory (US) [view abstract]
Segment Anything Model (SAM) has shown unprecedented generalization for open-world segmentation, yet it remains computationally expensive and poorly adapted to medical domains. We introduce Med-MoSAM, an efficient and adaptive framework that integrates a Mixture-of-Experts (MoE) design within SAM s architecture. Instead of full fine-tuning, Med-MoSAM inserts sparse expert adapters into frozen SAM layers and uses a lesion-aware routing gate to dynamically activate a small subset of experts depending on the anatomical structure or modality. This design preserves SAM s general representation while enabling domain-specific specialization. Experiments across multiple 2D and 3D medical segmentation benchmarks including AMOS, LiTS, and BraTS demonstrate that Med-MoSAM reduces computation by over 60% while achieving superior Dice and HD95 performance compared to MedSAM and LoRA-SAM. The approach offers a scalable path toward foundation-model efficiency and generalization in clinical imaging.
08:50HPCI-196
Focus: Fused observation of channels for unveiling spectra, [view abstract]
Hyperspectral imaging (HSI) captures hundreds of narrow, contiguous wavelength bands, making it a powerful tool in biology, agriculture, and environmental monitoring. However, interpreting Vision Transformers (ViTs) in this setting remains largely unexplored due to two key challenges: (1) existing saliency methods struggle to capture meaningful spectral cues, often collapsing attention onto the class token, and (2) full-spectrum ViTs are computationally prohibitive for interpretability, given the high-dimensional nature of HSI data. We present FOCUS, the first framework that enables reliable and efficient spatial-spectral interpretability for frozen ViTs. FOCUS introduces two core components: class-specific spectral prompts that guide attention toward semantically meaningful wavelength groups, and a learnable [SINK] token trained with an attraction loss to absorb noisy or redundant attention. Together, these designs make it possible to generate stable and interpretable 3D saliency maps and spectral importance curves in a single forward pass, without any gradient backpropagation or backbone modification. FOCUS improves band-level IoU by 15 percent, reduces attention collapse by over 40 percent, and produces saliency results that align closely with expert annotations. With less than 1 percent parameter overhead, our method makes high-resolution ViT interpretability practical for real-world hyperspectral applications, bridging a long-standing gap between black-box modeling and trustworthy HSI decision-making.
09:10HPCI-197
ORBIT-2: Exascale vision transformer for weather and climate downscaling, [view abstract]
Sparse observations and coarse-resolution climate models limit effective regional decision-making, underscoring the need for robust downscaling. However, existing AI methods struggle with generalization across variables and geographies and are constrained by the quadratic complexity of Vision Transformer (ViT) self-attention. We introduce ORBIT-2, a scalable foundation model for global, hyper-resolution climate downscaling. ORBIT-2 incorporates two key innovations: (1) Residual Slim ViT (Reslim), a lightweight architecture with residual learning and Bayesian regularization for efficient, robust prediction; and (2) TILES, a tile-wise sequence scaling algorithm that reduces self-attention complexity from quadratic to linear, enabling long-sequence processing and massive parallelism. ORBIT-2 scales to 10 billion parameters across 65,536 GPUs, achieving up to 4.1 exaFLOPS sustained throughput and 74--98% strong scaling efficiency. It supports downscaling to 0.9 km global resolution and processes sequences up to 4.2 billion tokens. On 7 km resolution benchmarks, ORBIT-2 achieves high accuracy with R^2 scores in the range of 0.98--0.99 against observational data.
09:30HPCI-198
HistoWAS: A novel pathomics framework for large-scale feature-wide association studies of tissue topology and patient outcomes, Yuechen Yang, Vanderbilt University (US); Junlin Guo, Vanderbilt University (US); Yanfan Zhu, Vanderbilt University (US); Jialin Yue, Washington University in St. Louis (US); Haichun Yang, ; Yuankai Huo, Vanderbilt University (US) [view abstract]
The high-throughput extraction of "pathomic" features from Whole Slide Images (WSIs) in digital pathology presents new opportunities but also introduces the "curse of dimensionality." Current approaches often rely on complex, "black-box" models that limit interpretability and hinder the discovery of novel biomarkers. To address this, we introduce HistoWAS (Histology-Wide Association Study), a computational framework for systematically screening pathomic features for significant associations with clinical endpoints. The core innovations of HistoWAS are twofold: 1) an expanded feature space that moves beyond conventional single-object metrics to include novel topological and spatial features, adapted from Geographic Information Systems (GIS) point pattern analysis (e.g., Ripley's K/L-function), to quantify tissue micro-architecture; and 2) a scalable association study engine, inspired by phenome-wide association study(PheWAS), that performs mass univariate regression for each feature, incorporating rigorous corrections for multiple hypothesis testing. In a validation study on a renal interstitial fibrosis cohort from the Kidney Precision Medicine Project (KPMP), comprising 236 WSIs from 103 patients, we applied HistoWAS to analyze 112 pathomic features spanning four categories. Through quantitative assessment of WSI, the analytical framework successfully identifies key indicators correlated with pathological states. By utilizing intuitive visualizations, it reveals novel biomarkers that might be obscured in complex models, offering a robust paradigm for exploratory research in pathology.
Efficient Vision AI Model II
11:00 - 12:20
Grand Peninsula F
11:00HPCI-199
HoloHisto: End-to-end gigapixel WSI segmentation with 4K resolution sequential tokenization, Yucheng Tang, NVIDIA Corp. (US); Yufan He, NVIDIA Corp. (US); Vishwesh Nath, NVIDIA Corp.; Pengfeig Guo, NVIDIA Corp.; Ruining Deng, Vanderbilt University; Tianyuan Yao, Vanderbilt University; Quan Liu, Vanderbilt University; Can Cui, Vanderbilt University; Yuechen Yang, Vanderbilt University; Mengmeng Yin, Vanderbilt University; Ziyue Xu, NVIDIA Corp.; Holger Roth, NVIDIA Corp.; Daguang Xu, NVIDIA Corp. [view abstract]
11:20HPCI-200
Scaling transformers for spatio-temporal data, Alexander Kiefer, University of Tennessee (US); Shashank Subramanian, NERSC (US); Dmitriy Morozov, Lawrence Berkeley National Laboratory (US); Romain , Oak Ridge National Laboratory (US); Xiao Wang, Oak Ridge National Laboratory (US); Prasanna Balaprakash, Oak Ridge National Laboratory (US) [view abstract]
The introduction of the transformer architecture and subsequent adoption by the NLP and CV communities have led to the widespread creation of large-scale foundation models. The fundamental principle behind foundation models is the ability to consistently extract greater performance with increasing model sizes. This, however, proves challenging when applied to spatio-temporal data, where the dimensionality of data grows and relationships between dimensions become more complex. We explore a novel model architecture designed for efficient scaling on spatio-temporal data and its scaling characteristics through a systematic scaling study on the ERA5 dataset. We show that through the use of spatial and tensor parallelism, we can efficiently scale to over one billion parameters.
11:40HPCI-201
MAGIC at scale: LoRA-efficient, expert-guided diffusion for clinically accurate dermatology image synthesis, Janet Wang, Tulane University (US); Yunbei Zhang, Tulane University (US); Zhengming Ding, Tulane University (US); Jihun Hamm, Tulane University (US); Xiao Wang, Oak Ridge National Laboratory (US) [view abstract]
We present MAGIC, a scalable, LoRA-efficient framework for training text-to-image diffusion models to synthesize clinically accurate dermatology images. Unlike prior approaches that rely on post-hoc filtering, MAGIC integrates expert-crafted clinical checklists during training using a multimodal LLM (MLLM) to score images and convert criteria into pairwise preferences for Direct Preference Optimization (DPO). To improve runtime and hardware efficiency, we extend the pipeline with Low-Rank Adaptation (LoRA) adapters on UNet attention blocks, preserving base weights while reducing trainable parameters and memory. MAGIC offers a general recipe for high-performance human-in-the-loop alignment of generative models in safety-critical imaging domains.
12:00HPCI-202
Uncertainty-driven efficiency in vision transformers, [view abstract]
Vision Transformers (ViTs) face a critical computational bottleneck due to the quadratic complexity of the self-attention mechanism. While dynamic token pruning offers significant acceleration, current approaches prioritize In-Distribution (ID) accuracy, creating a "Robustness Gap" where Out-of-Distribution (OOD) anomalies are frequently discarded as redundant. This paper introduces Evidential Token Pruning (ETP), a framework designed to reconcile efficiency with reliability. To ensure wall-clock acceleration, we reject computationally expensive ensembles in favor of single-pass deterministic uncertainty methods. We theoretically decompose token utility into aleatoric and epistemic components, establishing a policy where high-epistemic tokens (anomalies) are preserved while aleatoric noise is pruned. The proposed Evidential Pruning Module (EPM) employs spectral normalization and a bivariate selection logic, effectively bounding information loss by retaining tokens based on a composite of learned saliency and epistemic uncertainty. This approach minimizes FLOPs while immunizing the model against the "tunnel vision" typical of standard pruning methods.