multilayer perceptron
PulseAugur coverage of multilayer perceptron — every cluster mentioning multilayer perceptron across labs, papers, and developer communities, ranked by signal.
6 天有情绪数据
-
New mechanistic estimation method outperforms sampling for wide random MLPs
Researchers have developed a new method for estimating the expected output of wide, randomly initialized multilayer perceptrons (MLPs) without needing to run samples through the model. This "mechanistic estimation" appr…
-
TabSurv adapts tabular neural networks for improved survival analysis
Researchers have introduced TabSurv, a novel approach that adapts modern tabular neural network architectures for survival analysis tasks. This method utilizes a new histogram loss function called SurvHL, which is desig…
-
Manokhin 概率矩阵为分类器质量提供新框架
研究人员引入了 Manokhin 概率矩阵，这是一个旨在评估分类器概率预测质量的新诊断框架。该框架区分了可靠性和分辨率，将分类器分为四种原型：Eagle、Bull、Sloth 和 Mole。一项对 21 个分类器和 30 个任务进行的实证研究发现，像 CatBoost 和 Random Forest 这样的模型是 Eagles，而 XGBoost 和 LightGBM 是 Bulls，这对事后校准具有特定意义。
-
New framework evaluates autonomous driving AI robustness against real-world adversarial attacks
Researchers have developed a new framework for evaluating the real-time robustness of autonomous driving systems against adversarial attacks. This approach utilizes real-world intersection driving data, moving beyond pu…
-
联邦学习框架以97%的准确率增强5G干扰检测能力
研究人员开发了一个联邦学习框架，用于检测5G网络中的射频（RF）干扰攻击。该方法使用同步信号块（SSB）的同相（In-phase）和正交（Quadrature）样本来训练一维卷积神经网络（CNN），从而在不共享原始信号数据的情况下，实现用户设备之间的协作模型训练。联邦学习方法达到了97%的准确率和F1分数，在保护用户隐私的同时，性能优于集中式机器学习模型。
-
Researchers develop SNMF for interpretable LLM feature analysis
Researchers have developed a new method for understanding the internal workings of large language models by decomposing MLP activations. This technique, semi-nonnegative matrix factorization (SNMF), identifies interpret…
-
P1-KAN network offers improved accuracy and convergence over MLPs
Researchers have introduced P1-KAN, a novel Kolmogorov-Arnold Network designed to approximate complex, irregular functions in high-dimensional spaces. The paper provides theoretical error bounds and universal approximat…
-
Researchers propose linear-time global visual modeling by replacing attention with dynamic parameterization.
Researchers have developed a new method for visual modeling that achieves global sequence modeling capabilities without relying on explicit attention mechanisms. By reframing attention as a Multi-Layer Perceptron with d…
-
MSMixer model enhances long-term time series forecasting with multi-scale temporal mixing
Researchers have introduced MSMixer, a novel multi-scale MLP architecture designed for long-term time series forecasting. This model simultaneously processes data at different temporal resolutions (1x, 4x, and 16x) usin…
-
AI reconstructs high-resolution diffusion MRI from single views, accelerating scans
Researchers have developed a self-supervised Spatial-Angular Implicit Neural Representation (SA-INR) to reconstruct high-resolution diffusion MRI (dMRI) from fewer rotating views. This method, an MLP conditioned on stru…
-
研究人员发现随机删除数据可改进自适应强化学习策略
研究人员发现，随机删除一部分训练数据可以显著提高自适应强化学习策略的性能。这种反直觉的技术通过隐式地降低来自与部署环境不同分布的旧数据的权重来提供帮助。该方法将某些网络架构的鲁棒性差距最多降低了30%，并能使较小的模型在没有删除的情况下优于训练得更大的模型。理论分析表明，当训练和部署分布不匹配时，尤其是在中等正则化和低信噪比的情况下，这种删除策略是有益的。
-
新AI方法提升时间序列预测的准确性和可解释性
研究人员引入了几种新的时间序列预测方法，旨在提高准确性和泛化能力。MeLISA是一种无潜在变量的自回归模型，可提高回溯效率和长视界统计准确性。Temporal Functional Circuits利用Kolmogorov-Arnold Networks (KANs)为预测提供忠实且与时间相关的解释。Dynamic Pattern Recalibration (DPR)提供了一种与骨干网络无关的令牌级重新校准机制，以适应不断变化的局部…
-
ITS-Mina framework offers competitive multivariate time series forecasting with MLPs
Researchers have introduced ITS-Mina, a new framework for multivariate time series forecasting that utilizes a simpler MLP-based architecture. This approach incorporates an iterative refinement mechanism to deepen model…
-
DPN-LE方法以最小的神经元干预精确编辑LLM个性
研究人员开发了DPN-LE，一种通过靶向特定神经元来编辑大型语言模型“个性”的新颖方法。现有技术通常通过修改过多神经元（其中许多是多功能的）来降低整体模型性能。DPN-LE通过对比MLP激活来识别特定于个性的神经元，并使用双重标准过滤方法来分离相关的神经元子集。该方法仅干预一小部分神经元，在保持通用能力的同时实现精确的个性控制。
-
IDOBE benchmark ecosystem offers standardized evaluation for outbreak forecasting models
Researchers have introduced IDOBE, a new benchmark ecosystem designed to evaluate infectious disease outbreak forecasting models. This curated collection includes over 10,000 outbreaks derived from epidemiological time …
-
Deep Transformer models show synchronization by noise in new research
Researchers have published a paper detailing the mathematical behavior of deep transformer models. The study proves that the layerwise evolution of tokens within these models converges to a continuous-time stochastic in…
-
Robotic fruit picking sensors analyzed for improved success rates
Researchers have developed a multimodal sensing suite for robotic fruit harvesting to improve pick success detection. The system analyzes which sensors are most informative during different stages of the picking process…
-
新框架使用同态加密在加密数据上训练机器学习模型
研究人员开发了一个使用同态加密在加密数据上训练机器学习模型的隐私保护框架。这种方法允许在加密数据上进行计算，在整个机器学习过程中保护敏感信息。该框架成功演示了 K-近邻和线性回归模型的训练，取得了与在未加密数据上训练的模型相当的性能，尽管计算开销和噪声管理方面仍存在挑战。
-
New Graph Transformer models improve microservice tail latency prediction
Two new research papers propose advanced methods for predicting tail latency in microservice systems. The first, STLGT, uses a graph transformer to model service dependencies and a temporal module for workload dynamics,…
-
MLP 跳跃连接无法被吸收进无残差模型
研究人员调查了一个单隐藏层 MLP 周围的跳跃连接是否可以被吸收进一个相同宽度的无残差 MLP。他们发现，对于 ReLU^2 和 ReGLU 等某些激活函数，由于次数参数的原因，吸收是不可能的。对于 SwiGLU 和 GeGLU 等门控激活函数，线性化参数也得出了相同的结论。虽然在特定的、非通用的权重条件下，吸收对于无门控的 ReLU 和 GELU 是可能的，但跳跃连接和无残差的 MLP 通常代表不同的函数类别。