Deep Neural Networks
PulseAugur coverage of Deep Neural Networks — every cluster mentioning Deep Neural Networks across labs, papers, and developer communities, ranked by signal.
6 天有情绪数据
-
AI在成像中的幻觉与逆问题极限相关
研究人员开发了一个理论框架，用于理解和量化用于逆问题（如医学成像）的AI模型中的“幻觉”。研究表明，这些逼真但错误的细节可能源于问题本身固有的病态性质，而不仅仅是特定模型。新方法提供了幻觉幅度的可计算界限以及评估重建忠实度的算法，证明了其在各种成像任务和现代生成模型中的广泛适用性。
-
Study evaluates transfer learning for deep neural networks in image classification
Researchers explored how to best select pre-trained deep neural networks for image classification tasks. They adapted eleven models, originally trained on ImageNet, to five distinct target datasets. The study evaluated …
-
Neuromorphic depth estimation uses event cameras with uncertainty modeling
Researchers have developed a neuromorphic approach to monocular depth estimation using event cameras, which offer advantages like high temporal resolution and dynamic range. Their deep neural network models predict per-…
-
New methods improve Laplace approximation for neural network uncertainty
Researchers have developed new methods for approximating the Laplace approximation in deep neural networks, addressing the computational challenges of inverting large Hessian matrices. The proposed Gradient-Laplace and …
-
PoTAcc流水线在边缘设备上加速二的幂量化深度神经网络
研究人员开发了PoTAcc，一个开源流水线，旨在加速资源受限的边缘设备上二的幂（PoT）量化深度神经网络（DNN）的部署。该系统通过TensorFlow Lite促进这些模型的准备和部署，支持仅CPU配置以及带有定制加速器的混合CPU-FPGA系统。评估表明，使用PoTAcc的CPU-加速器设计在特定FPGA板上与仅CPU执行相比，实现了高达3.6倍的速度提升和78%的能耗降低。
-
New dynamic self-optimizing control method uses deep neural networks
Researchers have introduced a new framework for dynamic self-optimizing control, extending the concept to dynamic processes beyond steady-state applications. The paper proposes "dynamic controlled variables" (DCVs) and …
-
新的AdaLoc方法确保了可适应的AI模型使用控制
研究人员开发了一种名为AdaLoc的新方法，通过将访问密钥嵌入到模型参数的子集中来增强深度神经网络（DNN）的安全性。这种方法实现了可适应的模型使用控制，这意味着即使在微调或特定任务更新后，也可以在不进行完全重新密钥设置的情况下，将模型的效用恢复到授权状态。在各种基准测试和架构上的实验表明，AdaLoc在为授权用户保持高精度的同时，能够显著降低未经授权访问的性能，使其下降到接近随机猜测的水平。
-
Researchers explore neural network complexity, computation, and graph theory connections
Researchers are exploring new theoretical frameworks and computational models for neural networks. One paper introduces a unified framework to analyze and construct deep neural networks by modeling tensor operations, re…
-
Researchers develop BadSNN to exploit spiking neuron hyperparameters for backdoor attacks
Researchers have developed "BadSNN," a novel backdoor attack targeting Spiking Neural Networks (SNNs). This attack exploits variations in the hyperparameters of spiking neurons, such as the Leaky Integrate-and-Fire mode…
-
Deep Neural Networks Achieve Universality via Lindeberg Exchange Principle
Researchers have developed a new approach to understand the behavior of deep neural networks in their infinite-width limit. By applying a Lindeberg principle specifically adapted for deep neural networks, they can quant…
-
Researchers develop PyFair framework for testing neural network individual fairness
Researchers have developed PyFair, a new framework designed to formally assess and verify individual fairness in deep neural networks. This system adapts concolic testing techniques to systematically explore network beh…
-
New 'sphere cloud' method enhances privacy in 3D visual localization
Researchers have developed a novel privacy-preserving technique for visual localization using a "sphere cloud" representation. This method addresses concerns about deep neural networks reconstructing private maps from 3…
-
新的VCON框架可实现平滑迭代的深度神经网络压缩，准确率损失极小
研究人员推出了一种名为消失的贡献（VCON）的新型框架，旨在简化深度神经网络的压缩过程。VCON通过在微调期间并行运行原始模型和压缩模型，实现了向压缩模型的更平滑、迭代的过渡。这种方法逐渐减少了未压缩模型的影响，同时增加了压缩模型的贡献，从而提高了稳定性和降低了准确率损失。在计算机视觉和自然语言处理任务上的评估表明，VCON持续提高了性能，典型准确率提升超过1%，某些配置的提升超过15%。
-
深度神经网络可证明地克服了偏微分方程的维度灾难
研究人员证明了深度神经网络（DNN）在逼近Kolmogorov偏微分方程解时可以克服维度灾难。这项数学证明扩展了先前的发现，表明使用ReLU、Leaky ReLU和Softplus激活函数的网络可以在不导致计算成本相对于问题维度呈指数级增长的情况下，实现逼近精度。该工作在$L^p$意义下，针对广泛的$p$值证明了这种能力。
-
AI paradigm discovers explainable scientific equations, outperforming deep learning
Researchers have introduced a new paradigm called machine collective intelligence, designed to autonomously discover governing equations from empirical data. This approach combines symbolic reasoning with metaheuristics…
-
DNNs mitigate dimensional collapse in feature interaction models, study finds
This paper investigates the role of Deep Neural Networks (DNNs) in feature interaction recommendation models, addressing a debate on their ability to capture complex interactions. The research proposes a new perspective…
-
DNNs, Dataset Statistics, and Correlation Functions
A new paper proposes that the success of deep neural networks (DNNs) in image recognition tasks stems from their ability to discover high-order correlation functions within datasets. The authors argue that DNNs effectiv…
-
Researchers explore new methods for detecting adversarial data and analyzing active learning algorithms
Researchers have developed a new method to detect adversarial data in deep neural networks by formally proving an adversarial noise amplification theorem. This theoretical framework underpins a novel training methodolog…