residual neural network
PulseAugur coverage of residual neural network — every cluster mentioning residual neural network across labs, papers, and developer communities, ranked by signal.
7 天有情绪数据
-
新的RBDC协议将视觉模型训练成本降低了30%
研究人员开发了一种名为RBDC的新训练协议，以提高训练大型视觉模型的可资源效率。该方法通过无参数的块对角线方式递归地耦合独立训练的、更窄的模型。在ImageNet上使用Vision Transformers和ResNets进行的评估表明，与现有的增长方法相比，FLOPs减少了30%，准确率相当，并且在相同的训练FLOPs下性能有所提高。RBDC训练的模型在作为对象检测和实例分割等下游任务的骨干网络方面也显示出增强的效用。
-
新的MVProbe框架通过权重空间学习分析AI模型
研究人员开发了MVProbe，一种新颖的多视图探测框架，旨在直接从其参数分析大型开源AI模型。该方法通过提取可学习的探测向量的表示来解决处理完整模型权重的计算限制。MVProbe通过纳入高阶相关性模式，增强了现有的单视图探测技术，在ResNet和Stable Diffusion LoRA适配器等各种架构的模型丛林基准测试中表现优于先前的方法。
-
已识别出神经网络权重漂移是训练动态问题
研究人员在神经网络中发现了一种称为“权重漂移”的现象，其中优化过程会无意中将权重推向负值。这种漂移独立于训练数据，在使用标准损失函数和 ReLU、GELU 等常见激活函数时会出现。研究表明，这种漂移会导致显著的激活稀疏性，可能影响模型准确性，并且还会放大 Transformer 层中的激活尖峰。
-
FAIR-Pruner框架支持自适应逐层神经网络剪枝
研究人员开发了FAIR-Pruner，一个用于深度神经网络自动、逐层结构化剪枝的新框架。该方法通过使用移除导向和保护导向的信号，自适应地在网络层之间分配稀疏度。在包括视觉模型和Qwen1.5-MoE模型在内的各种数据集和模型架构上的实验表明，FAIR-Pruner实现了强大的精度-压缩权衡。该框架可作为一个开源包使用。
-
深度学习模型在COVID-19图像分类中达到98%的准确率
研究人员对用于从CT和X射线肺部影像中分类COVID-19的各种深度学习架构进行了综合比较。该研究使用了包括VGG、Densenet、Resnet、MobileNet、Xception、EfficientNet和NasNet在内的预训练模型。结果表明，Resnet和VGG架构在区分COVID-19阳性病例与健康肺部方面达到了95%至98%的高准确率，优于以往的文献发现。
-
MambaGaze 框架使用 Mamba-2 进行认知负荷评估
研究人员开发了 MambaGaze，一个利用眼动追踪数据准确评估认知负荷的新框架。该系统利用双向 Mamba-2 有效建模长程时间依赖性，并采用 XMD 编码方法显式处理因眨眼等原因造成的缺失数据。MambaGaze 在基准数据集上的表现优于现有模型，并可在 NVIDIA Jetson 平台等边缘设备上进行实时部署。
-
StableGrad stabilizes deep neural network training without batch normalization
Researchers have introduced StableGrad, a novel optimizer-level mechanism designed to control the scale of activations and gradients in deep neural networks. This method aims to prevent training instability without rely…
-
New neural operator frameworks promise enhanced scientific computing
Researchers are exploring advanced neural operator frameworks to enhance scientific computing. One paper introduces the Infinite-order Kernel Neural Operator (IKNO), which uses infinite-order kernel integrals for improv…
-
CNN architecture evolution driven by depth, scaling, and training recipes
A recent analysis delves into the evolution of Convolutional Neural Network (CNN) architectures, specifically examining ResNet, EfficientNet, and ConvNeXt. The author investigates whether advancements in state-of-the-ar…
-
New theory reveals optimal learning rate schedules for deep learning
Researchers have developed a theoretical framework for optimal learning rate schedules in deep learning, specifically analyzing a random feature model trained with stochastic gradient descent. The study identifies two d…
-
CircleID competition sets new benchmark for writer ID from circles
A new competition, CircleID, has been launched for the ICDAR 2026 competition, focusing on writer identification and pen classification using only scanned hand-drawn circles. The dataset includes over 46,000 circle imag…
-
New models and datasets advance egocentric hand pose forecasting
Researchers have introduced EggHand, a new multimodal foundation model designed for egocentric hand pose forecasting from video. This model integrates semantic reasoning with dynamic motion modeling, utilizing a Vision-…
-
Contact Wasserstein Geodesics offer new approach to Schrödinger Bridges
Researchers have developed a novel reformulation of the Schrödinger Bridge problem, termed the non-conservative generalized Schrödinger bridge (NCGSB). This new approach overcomes limitations of previous methods by allo…
-
Evolutionary fine tuning boosts accuracy of quantized deep learning models
Researchers have developed a novel method for improving the accuracy of quantized deep learning models by employing an evolutionary strategy. This approach fine-tunes pre-trained and quantized models by iteratively adju…
-
New AI framework learns classification losses without real data
Researchers have developed a new framework called Evolutionary Dynamic Loss (EDL) for pretraining classification losses without using real data. EDL learns a transferable loss function by generating synthetic prediction…
-
研究人员开发用于无分布预训练的进化动态损失
研究人员开发了一个名为进化动态损失（EDL）的新框架，用于预训练分类损失。EDL使用合成数据学习可迁移的损失函数，避免了在主要预训练阶段需要真实样本。该框架通过进化策略将损失优化为一个轻量级网络，并结合混沌变异来增强探索和改善收敛性。在CIFAR-10上的实验表明，EDL可以有效地替代交叉熵并达到相当或更好的准确率。
-
MSACT improves robot fine manipulation with stable, low-latency spatial alignment
Researchers have developed MSACT, a novel method for improving fine manipulation in robotics, particularly for bimanual tasks. This approach uses a multistage spatial attention module to extract stable 2D attention poin…
-
DeepWeightFlow generates diverse, high-accuracy neural network weights efficiently
Researchers have introduced DeepWeightFlow, a novel generative model designed to create neural network weights directly in weight space. This approach addresses challenges with high-dimensional weight spaces and network…
-
AI segmentation study highlights PE detection challenges, offers open-weight model
Researchers have identified significant limitations in current pulmonary embolism (PE) segmentation algorithms, citing issues with small datasets, lack of reproducibility, and insufficient comparative evaluations. Their…
-
100,000 Yuan Investment: Latest Interview with Princeton's Zhuang Liu: Architecture Isn't That Important, Data is King
Princeton Assistant Professor Liu Zhuang argues that AI architecture is less critical than previously thought, with data scale and diversity being the primary drivers of progress. In a recent interview, he highlighted t…