VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model

Medical image segmentation involves partitioning medical images into meaningful regions, with a focus on identifying anatomical structures and lesions. It has broad applications in healthcare, and deep learning methods have enabled significant advancements in automating this process. Recently, the introduction of the Segmentation Anything Model (SAM), the first foundation model for segmentation task, has prompted researchers to adapt it for the medical domain to improve performance across various tasks. However, SAM’s large model size and high GPU requirements hinder its scalability and development in the medical domain. To address these challenges, research has increasingly focused on lightweight adaptations of SAM to reduce its parameter count, enabling training with limited GPU resources while maintaining competitive segmentation performance. In this work, we propose MCP-MedSAM, a powerful and lightweight medical SAM model designed to be trainable on a single A100 GPU with 40GB of memory within one day while delivering superior segmentation performance. Recognizing the significant internal differences between modalities and the need for direct segmentation target information within bounding boxes, we introduce two kinds of prompts: the modality prompt and the content prompt. After passing through the prompt encoder, their embedding representations can further improve the segmentation performance by incorporating more relevant information without adding significant training overhead. Additionally, we adopt an effective modality-based data sampling strategy to address data imbalance between modalities, ensuring more balanced performance across all modalities. Our method was trained and evaluated using a large-scale challenge dataset, compared to top-ranking methods on the challenge leaderboard, MCP-MedSAM achieved superior performance while requiring only one day of training on a single GPU. The code is publicly available at <a href='https://github.com/dong845/MCP-MedSAM'>https://github.com/dong845/MCP-MedSAM</a>

In the era of Big Data, large-scale machine learning models have revolutionized various fields, driving significant advancements. However, large-scale model training demands high financial and computational resources, which are only affordable by a few technological giants and well-funded institutions. In this case, common users like mobile users, the real creators of valuable data, are often excluded from fully benefiting due to the barriers, while the current methods for accessing largescale models either limit user ownership or lack sustainability. This growing gap highlights the urgent need for a collaborative model training approach, allowing common users to train and share models. However, existing collaborative model training paradigms, especially federated learning (FL), primarily focus on data privacy and group-based model aggregation. To this end, this paper intends to address this issue by proposing a novel training paradigm named decentralized relay learning (DeRelayL), a sustainable learning system where permissionless participants can contribute to model training in a relay-like manner and share the model. In detail, this paper presents the architecture and workflow of DeRelayL, designs incentive mechanisms to ensure sustainability, and conducts theoretical analysis and numerical simulations to demonstrate its effectiveness

Integrating low-level edge features has been proven to be effective in preserving clear boundaries of salient objects. However, the locality of edge features makes it difficult to capture globally salient edges, leading to distraction in the final predictions. To address this problem, we propose to produce distraction-free edge features by incorporating cross-scale holistic interdependencies between high-level features. In particular, we first formulate our edge features extraction process as a boundary-filling problem. In this way, we enforce edge features to focus on closed boundaries instead of those disconnected background edges. Secondly, we propose to explore cross-scale holistic contextual connections between every position pair of high-level feature maps regardless of their distances across different scales. It selectively aggregates features at each position based on its connections to all the others, simulating the “contrast” stimulus of visual saliency. Finally, we present a complementary features integration module to fuse low- and high-level features according to their properties. Experimental results demonstrate our proposed method outperforms previous state-of-the-art methods on the benchmark datasets, with the fast inference speed of 30 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FPS</i> on a single GPU.

Vision Transformers (ViTs) have emerged as state-of-the-art models for various vision tasks recently. However, their heavy computation costs remain daunting for resource-limited devices. To address this, researchers have dedicated themselves to compressing redundant information in ViTs for acceleration. However, existing approaches generally sparsely drop redundant image tokens by token pruning or brutally remove channels by channel pruning, leading to a sub-optimal balance between model performance and inference speed. Moreover, they struggle when transferring compressed models to downstream vision tasks that require the spatial structure of images, such as semantic segmentation. To tackle these issues, we propose CAIT, a joint compression method for ViTs that achieves a harmonious blend of high accuracy, fast inference speed, and favorable transferability to downstream tasks. Specifically, we introduce an asymmetric token merging (ATME) strategy to effectively integrate neighboring tokens. It can successfully compress redundant token information while preserving the spatial structure of images. On top of it, we further design a consistent dynamic channel pruning (CDCP) strategy to dynamically prune unimportant channels in ViTs. Thanks to CDCP, insignificant channels in multi-head self-attention modules of ViTs can be pruned uniformly, significantly enhancing the model compression. Extensive experiments on multiple benchmark datasets show that our proposed method can achieve state-of-the-art performance across various ViTs.

In early design phases and during software evolution, design-time energy efficiency analyses enable software architects to reason on the effect of design decisions on energy efficiency. Energy efficiency analyses rely on accurate power models to estimate power consumption. Deriving power models that are both accurate and usable for design time predictions requires extensive measurements and manual analysis. Existing approaches that aim to automate the extraction of power models focus on the construction of models for runtime estimation of power consumption. Power models constructed by these approaches do not allow users to identify the central set of system metrics that impact energy efficiency prediction accuracy. The identification of these central metrics is important for design time analyses, as an accurate prediction of each metric incurs modeling effort. We propose a methodology for the automated construction of multi-metric power models using systematic experimentation. Our approach enables the automated training and selection of power models for the design time prediction of power consumption. We validate our approach by evaluating the prediction accuracy of derived power models for a set of enterprise and data-intensive application benchmarks.

To achieve high accuracy when performing deep learning, it is necessary to use a large-scale training model. However, due to the limitations of GPU memory, it is difficult to train large-scale training models within a single GPU. NVIDIA introduced a technology called CUDA Unified Memory with CUDA 6 to overcome the limitations of GPU memory by virtually combining GPU memory and CPU memory. In addition, in CUDA 8, memory advise options are introduced to efficiently utilize CUDA Unified Memory. In this work, we propose a newly optimized scheme based on CUDA Unified Memory to efficiently use GPU memory by applying different memory advise to each data type according to access patterns in deep learning training. We apply CUDA Unified Memory technology to PyTorch to see the performance of large-scale learning models through the expanded GPU memory. We conduct comprehensive experiments on how to efficiently utilize Unified Memory by applying memory advises when performing deep learning. As a result, when the data used for deep learning are divided into three types and a memory advise is applied to the data according to the access pattern, the deep learning execution time is reduced by 9.4% compared to the default Unified Memory.

Optimizing makespan and resource utilization for multi-DNN training in GPU cluster

In this work, we propose a framework for singleview hand mesh reconstruction, which can simultaneously achieve high reconstruction accuracy, fast inference speed, and temporal coherence. Specifically, for 2D encoding, we propose lightweight yet effective stacked structures. Regarding 3D decoding, we provide an efficient graph operator, namely depth-separable spiral convolution. Moreover, we present a novel feature lifting module for bridging the gap between 2D and 3D representations. This module begins with a map-based position regression (MapReg) block to integrate the merits of both heatmap encoding and position regression paradigms for improved 2D accuracy and temporal coherence. Furthermore, MapReg is followed by pose pooling and pose-to-vertex lifting approaches, which transform 2D pose encodings to semantic features of 3D vertices. Overall, our hand reconstruction framework, called MobRecon, comprises affordable computational costs and miniature model size, which reaches a high inference speed of 83FPS on Apple A14 CPU. Extensive experiments on popular datasets such as FreiHAND, RHD, and HO3Dv2 demonstrate that our MobRecon achieves superior performance on reconstruction accuracy and temporal coherence. Our code is publicly available at https://github.com/SeanChenxy/HandMesh.

Photoacoustic computed tomography (PACT) is a promising non-invasive imaging technique for both life science and clinical implementations. To achieve fast imaging speed, modern PACT systems have equipped arrays that have hundreds to thousands of ultrasound transducer (UST) elements, and the element number continues to increase. However, large number of UST elements with parallel data acquisition could generate a massive data size, making it very challenging to realize fast image reconstruction. Although several research groups have developed GPU-accelerated method for PACT, there lacks an explicit and feasible step-by-step description of GPU-based algorithms for various hardware platforms. In this study, we propose a comprehensive framework for developing GPU-accelerated PACT image reconstruction (GPU-accelerated photoacoustic computed tomography), to help the research community to grasp this advanced image reconstruction method. We leverage widely accessible open-source parallel computing tools, including Python multiprocessing-based parallelism, Taichi Lang for Python, CUDA, and possible other backends. We demonstrate that our framework promotes significant performance of PACT reconstruction, enabling faster analysis and real-time applications. Besides, we also described how to realize parallel computing on various hardware configurations, including multicore CPU, single GPU, and multiple GPUs platform. Notably, our framework can achieve an effective rate of times when reconstructing extremely large-scale three-dimensional PACT images on a dual-GPU platform compared to a 24-core workstation CPU. In this paper, we share example codes via GitHub. Our approach allows for easy adoption and adaptation by the research community, fostering implementations of PACT for both life science and medicine.

Multi-agent collaborative perception is expected to significantly improve perception performance by overcoming the limitations of single-agent perception through exchanging complementary information. However, training a robust collaborative perception model requires collecting sufficient training data that covers all possible collaboration scenarios, which is impractical due to intolerable deployment costs. Hence, the trained model is not robust against new traffic scenarios with inconsistent data distribution and fundamentally restricts its real-world applicability. Further, existing methods, such as domain adaptation, have mitigated this issue by exposing the deployment data during the training stage but incur a high training cost, which is infeasible for resource-constrained agents. In this paper, we propose a Parameter-Efficient Fine-Tuning-based lightweight framework, CoPEFT, for fast adapting a trained collaborative perception model to new deployment environments under low-cost conditions. CoPEFT develops a Collaboration Adapter and Agent Prompt to perform macro-level and micro-level adaptations separately. Specifically, the Collaboration Adapter utilizes the inherent knowledge from training data and limited deployment data to adapt the feature map to new data distribution. The Agent Prompt further enhances the Collaboration Adapter by inserting fine-grained contextual information about the environment. Extensive experiments demonstrate that our CoPEFT surpasses existing methods with less than 1\% trainable parameters, proving the effectiveness and efficiency of our proposed method.

Semantic segmentation is performed to understand an image at the pixel level; it is widely used in the field of autonomous driving. In recent years, deep neural networks achieve good accuracy performance; however, there exist few models that have a good trade-off between high accuracy and low inference time. In this paper, we propose a fast spatial feature network (FSFNet), an optimized lightweight semantic segmentation model using an accelerator, offering high performance as well as faster inference speed than current methods. FSFNet employs the FSF and MRA modules. The FSF module has three different types of subset modules to extract spatial features efficiently. They are designed in consideration of the size of the spatial domain. The multi-resolution aggregation module combines features that are extracted at different resolutions to reconstruct the segmentation image accurately. Our approach is able to run at over 203 FPS at full resolution ( $1024\times 2048$ ) in a single NVIDIA 1080Ti GPU, and obtains a result of 69.13% mIoU on the Cityscapes test dataset. Compared with existing models in real-time semantic segmentation, our proposed model retains remarkable accuracy while having high FPS that is over 30% faster than the state-of-the-art model. The experimental results proved that our model is an ideal approach for the Cityscapes dataset. The code is publicly available at: https://github.com/computervision8/FSFNet .

Training of large-scale deep learning models necessitates parallelizing the model and data across numerous devices, and the choice of parallelism strategy substantially depends on the training workloads such as memory consumption, computation cost, and communication cost. Current approaches generally assume uniform training workloads across samples in a given task. Thus, existing systems are designed to adopt a static parallelism strategy throughout one training process. Nevertheless, when training models with sequence inputs, this assumption fails due to the sequence length variation across samples. Consequently, training with a static parallelism strategy would result in sub-optimal performance.

Real-time and high-performance 3D object detection is an attractive research direction in autonomous driving. Recent studies prefer point based or voxel based convolution for achieving high performance. However, these methods suffer from the unsatisfied efficiency or complex customized convolution, making them unsuitable for applications with real-time requirements. In this paper, we present an efficient and effective 3D object detection framework, named RangeIoUDet that uses the range image as input. Benefiting from the dense representation of the range image, RangeIoUDet is entirely constructed based on 2D convolution, making it possible to have a fast inference speed. This model learns pointwise features from the range image, which is then passed to a region proposal network for predicting 3D bounding boxes. We optimize the pointwise feature and the 3D box via the point-based IoU and box-based IoU supervision, respectively. The point-based IoU supervision is proposed to make the network better learn the implicit 3D information encoded in the range image. The 3D Hybrid GIoU loss is introduced to generate high-quality boxes while providing an accurate quality evaluation. Through the point-based IoU and the box-based IoU, RangeIoUDet outperforms all single-stage models on the KITTI dataset, while running at 45 FPS for inference. Experiments on the self-built dataset further prove its effectiveness on different LIDAR sensors and object categories.

To address the growing size of AI model training data and the lack of a universal data selection methodology–factors that significantly drive up training costs–this paper presents the General Information Metrics Evaluation (GIME) method. GIME leverages general information metrics from Objective Information Theory (OIT), including volume, delay, scope, granularity, variety, duration, sampling rate, aggregation, coverage, distortion, and mismatch to optimize dataset selection for training purposes. Comprehensive experiments conducted across diverse domains, such as CTR Prediction, Civil Case Prediction, and Weather Forecasting, demonstrate that GIME effectively preserves model performance while substantially reducing both training time and costs. Additionally, applying GIME within the Judicial AI Program led to a remarkable 39.56% reduction in total model training expenses, underscoring its potential to support efficient and sustainable AI development.

In visual retrieval systems, updating the embedding model requires recomputing features for every piece of data. This expensive process is referred to as backfilling. Re-cently, the idea of backward compatible training (BCT) was proposed. To avoid the cost of backfilling, BCT modifies training of the new model to make its representations com-patible with those of the old model. However, BCT can sig-nificantly hinder the performance of the new model. In this work, we propose a new learning paradigm for representation learning: forward compatible training (FCT). In Fct, when the old model is trained, we also prepare for a future unknown version of the model. We propose learning side- information, an auxiliary feature for each sample which fa-cilitates future updates of the model. To develop a powerful and flexible frameworkfor model compatibility, we combine side-information with a forward transformation from old to newembeddings. Training of the new model is not modified, hence, its accuracy is not degraded. We demonstrate sig-nificant retrieval accuracy improvement compared to BCT for various datasets: ImageNet-1k (+18.1%), Places-365 (+5.4%), and VGG-Face2 (+8.3%). Fctobtains model compatibility when the new and old models are trained across different datasets, losses, and architectures. <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> Code available at https://github.com/apple/ml-fct.