Extends capacity analysis to neural networks, showing how capacity is distributed across layers.
problem How capacity is distributed in neural networks with non-linear layers.
method Introduces layer decoupling to quantify non-linear activation's impact, and uses a markovian rule for capacity propagation in deep networks.
result Shows that under certain conditions, capacity allocation in neural networks is equivalent to linear capacity allocation in an extended input space.
Backpropagation algorithm is indispensable for the training of feedforward neural networks. It requires propagating error gradients sequentially from the output layer all the way back to the input layer. The backward locking in backpropagation algorithm constrains us from updating network layers in parallel and fully l…
Study shows how large neural networks avoid overfitting through decoupling of feature learning and complexity growth.
problem Understanding inductive bias and generalization in large neural networks.
method Dynamical mean field theory applied to large two-layer networks.
result Training dynamics of large networks exhibit a separation of timescales, decoupling feature learning and overfitting.
A commonly cited inefficiency of neural network training by back-propagation is the update locking problem: each layer must wait for the signal to propagate through the full network before updating. Several alternatives that can alleviate this issue have been proposed. In this context, we consider a simpler, but more e…
Improved DNN calibration without sacrificing accuracy.
problem Poor calibration of over-parametrized DNNs in safety-critical applications.
method Decoupling feature extraction and classification layers, and applying Gaussian priors.
result Significant improvement in model calibration with minimal training cost.
Two graph auto-encoders decouple feature propagation from graph convolution layers.
problem Designing efficient graph auto-encoders with fixed receptive fields.
method L-GAE and L-VGAE using linear matrix computation before auto-encoder input.
result Comparable performance to VGAEs with smaller, simpler networks.
New smart contract mechanisms evade traditional AML systems by decoupling transaction roles.
problem Current AML systems fail to track economic value migration in composable smart contracts.
method Introduce PEB separation and state-mediated value migration to demonstrate how traditional tracing fails.
result Transfer-layer observation is incomplete and causally ambiguous in composable smart contracts.
This paper uncovers the low-rank structure of neural network Hessians.
problem Understanding the structure of Hessians in neural networks.
method Proposes a decoupling conjecture to decompose layer-wise Hessians into Kronecker products of smaller matrices.
result Proves the structure of top eigenspaces in 2-layer networks and shows high overlap in top eigenvectors across different models.
Regularized training of an autoencoder typically results in hidden unit biases that take on large negative values. We show that negative biases are a natural result of using a hidden layer whose responsibility is to both represent the input data and act as a selection mechanism that ensures sparsity of the representati…
Deep learning has gained great popularity due to its widespread success on many inference problems. We consider the application of deep learning to the sparse linear inverse problem encountered in compressive sensing, where one seeks to recover a sparse signal from a small number of noisy linear measurements. In this p…
NovoGrad improves deep learning training with adaptive moments and layer-wise normalization.
problem Training deep neural networks efficiently and effectively.
method Layer-wise adaptive moments with gradient normalization and decoupled weight decay.
result NovoGrad outperforms well-tuned SGD with momentum and Adam/AdamW in various tasks.
Unified SVD compression fails in practical tasks, highlighting the importance of per layer activation reconstruction.
problem The failure of a unified SVD compression method in practical tasks like perplexity and accuracy.
method Unified optimization problem for SVD based compression methods, focusing on cross-layer coupling.
result Downstream metrics like perplexity and accuracy degrade severely compared to standard per layer SVD LLM.
New heat semigroup characterizes Sobolev and BV spaces in Carnot groups.
problem Lack of explicit representations and symmetry in heat kernels in sub-Riemannian geometry.
method Establishes a new heat semigroup characterisation using integral decoupling property.
result Characterizes Sobolev and BV spaces in Carnot groups.
Sharp limits found for storing and retrieving input-output associations in linear associative memories.
problem Understanding the fundamental limits of storing and retrieving input-output associations in neural networks.
method Study of a minimal linear associative memory model, introducing a decoupled model and using statistical physics to characterize storage capacity.
result Linear associative memory can store up to 1/2 log(p) associations, providing a sharp statistical-physics characterization.
Inner product-based convolution has been a central component of convolutional neural networks (CNNs) and the key to learning visual representations. Inspired by the observation that CNN-learned features are naturally decoupled with the norm of features corresponding to the intra-class variation and the angle correspond…
FedMA improves federated learning for neural nets, matching and averaging model elements.
problem Federated learning's communication burden and model performance.
method Layer-wise construction of shared model by matching and averaging hidden elements.
result FedMA outperforms state-of-the-art algorithms on deep CNN and LSTM architectures.
Study on feature learning dynamics in infinite-depth neural networks, focusing on ResNets.
problem Understanding how features evolve during training in deep neural networks, especially in the large-depth limit.
method Conditional Gaussian representations and SDE system with decoupled backward weights.
result Depth-induced suppression of forward-backward coupling in infinite-depth networks, leading to a decoupled forward-backward SDE system.
Probabilistic decoupling separates labels from classes for improved classification.
problem Improving classification accuracy with noisy or partially labeled data.
method Probabilistic decoupling of labels from underlying classes.
result Method enhances performance on various classification tasks, including noisy and partially labeled data.
We investigate finite-time decoupled convergence in nonlinear two-time-scale stochastic approximation.
problem Achieving decoupled convergence in nonlinear two-time-scale stochastic approximation.
method Nested local linearity assumption, suitable step size selection, convergence analysis of matrix cross term, fourth-order moment convergence rates.
result Finite-time decoupled convergence rates can be achieved in nonlinear two-time-scale stochastic approximation with proper step size selection.
Deep Gaussian Processes (DGP) are hierarchical generalizations of Gaussian Processes (GP) that have proven to work effectively on a multiple supervised regression tasks. They combine the well calibrated uncertainty estimates of GPs with the great flexibility of multilayer models. In DGPs, given the inputs, the outputs …
Deep learning explained through spectral filtering of hierarchical features.
problem Understanding how deep neural networks learn useful representations from data.
method Neural Low-Degree Filtering (Neural LoFi) as a stylized limit of gradient-based training.
result Predicts how representations are selected layer by layer and explains emergence of concepts.
Spectral decoupling improves neural network generalization in medical imaging.
problem Poor generalization of neural networks trained on medical imaging data.
method Spectral decoupling, a regularization technique that encourages learning more features.
result Spectral decoupling increases network robustness and performance on external datasets.
A new neural network layer integrates graph learning into classification tasks.
problem Lack of relational information in standard deep learning architectures for label predictions.
method Derives backpropagation equations for a differentiable graph learning layer.
result Smooth label transitions, improved generalization, and robustness to adversarial attacks.
Decoupled GCN is shown to be equivalent to label propagation.
problem Improving semi-supervised node classification in graph learning.
method The paper proves the equivalence of decoupled GCN and label propagation, and proposes a new method named PTA.
result Decoupled GCN is equivalent to two-step label propagation and can automatically assign weights to pseudo-labels.
NVIDIA cuDNN is a low-level library that provides GPU kernels frequently used in deep learning. Specifically, cuDNN implements several equivalent convolution algorithms, whose performance and memory footprint may vary considerably, depending on the layer dimensions. When an algorithm is automatically selected by cuDNN,…
Artifical Neural Networks are a particular class of learning systems modeled after biological neural functions with an interesting penchant for Hebbian learning, that is "neurons that wire together, fire together". However, unlike their natural counterparts, artificial neural networks have a close and stringent couplin…
GraphSAINT improves GCN training efficiency and accuracy with graph sampling.
problem Neighbor explosion problem in minibatch training of GCNs.
method GraphSAINT constructs minibatches by sampling the training graph, ensuring fixed well-connected nodes in all layers.
result GraphSAINT achieves new state-of-the-art F1 scores for PPI and Reddit.
BPQP improves efficiency of differentiable optimization layers for deep learning.
problem Efficiency in differentiating optimization problems for deep learning models.
method Reformulates optimization problems as quadratic programming, enabling efficient gradient calculation.
result Significant improvement in efficiency (order of magnitude faster) compared to other differentiable optimization layers.
Many problems at the intersection of combinatorics and computer science require solving for a permutation that optimally matches, ranks, or sorts some data. These problems usually have a task-specific, often non-differentiable objective function that data-driven algorithms can use as a learning signal. In this paper, w…
The paper proposes a decoupled approach to efficiently estimate CoVaR, a measure of systemic financial risk.
problem Estimating CoVaR, a measure of systemic financial risk, is challenging due to zero-probability events and portfolio repricing.
method The paper introduces a decoupled approach using smoothing techniques and a functional perspective to model CoVaR.
result The decoupled estimator achieves a rate of convergence of approximately OmP(Γ−1/2). Improved sampling from mean-field stationary distributions.
problem Sampling from the stationary distribution of mean-field SDEs.
method Decoupling the problem into two aspects: approximation of mean-field SDE and sampling from finite-particle distribution.
result Improved guarantees in various settings, including optimizing neural networks.
This paper investigates the effectiveness of decoupled weight decay at the start of training.
problem The traditional approach to weight decay is not effective throughout training.
method The authors investigate decoupled weight decay, applying it only at the start of training.
result Applying weight decay only at the start of training stabilizes network weights and improves performance.
Clarifies method of phase synchronization for decoupling linear differential equations.
problem Velocity-dependent transformations in linear second-order differential equations.
method Linear transformation of coordinates and velocities.
result Velocity-dependent transformations do not preserve second-order character and define their own system.
New neural network models improve Granger Causality detection in non-linear systems.
problem Mischaracterization of Granger Causality in non-linear systems using traditional linear models.
method Proposes Learned Kernel VAR (LeKVAR) and decoupled penalties for GC estimation and lag selection.
result Improves GC detection in non-linear systems with computational efficiency.
Kahler geometry explains decoupling of Kerr perturbations.
problem Decoupling of curvature scalars in Kerr spacetime.
method Hidden Kahler structure in Kerr spacetime, showing decoupling as a consequence of Kahler geometry.
result Decoupling of Teukolsky equations on Kahler background.
New algorithms split deep learning tasks into representation and uncertainty estimation.
problem Challenges in uncertainty quantification for deep learning models.
method Proposes a two-stage approach: representation learning and uncertainty estimation.
result Simple methods outperform complex uncertainty layers in selective classification and out-of-distribution detection.
AdaDEM decouples EM into two parts to improve class overlap and uncertainty.
problem Improper EM limits its effectiveness in various machine learning tasks.
method Decouple EM into CADF and GMC, and AdaDEM normalizes CADF reward and uses MEC.
result AdaDEM outperforms classical EM and improves performance in noisy and dynamic environments.
Recent variants improve knowledge distillation performance.
problem Improving the performance of knowledge distillation.
method Introducing additional components or changing the learning process.
result These variants have shown promising results.
New algorithms for interpreting complex multivariate functions.
problem Hard interpretation of multivariate functions due to many parameters.
method Filtered tensor decompositions of derivative information.
result Nonparametric estimates of smooth decoupled functions.
STCN combines TCNs with stochastic latent variables for sequence modeling.
problem Performance gap between TCNs and stochastic RNNs, especially with multiple layers of random variables.
method Proposes a hierarchy of stochastic latent variables in a modular architecture.
result Achieves state-of-the-art log-likelihoods across various tasks.
New method combines spectral and sparse methods for Gaussian processes.
problem Efficiently fitting Gaussian processes to large datasets.
method Orthogonally decoupled variational Fourier features.
result Competitive performance on synthetic and real-world data.
We establish decoupled functional CLTs for two-time-scale stochastic approximation.
problem Understanding the asymptotic behavior of two-time-scale stochastic approximation.
method Martingale problem approach and auxiliary sequence.
result The limiting dynamics of two-time-scale SA are independent of each other.
Current reinforcement learning (RL) methods can successfully learn single tasks but often generalize poorly to modest perturbations in task domain or training procedure. In this work, we present a decoupled learning strategy for RL that creates a shared representation space where knowledge can be robustly transferred. …
Identifies interpretable generative model for multivariate data.
problem Black-box architectures of deep generative models are often unidentified and difficult to interpret.
method Introduces Deep Discrete Encoder (DDE) Copula, a hierarchical binary latent variable model inside a copula framework.
result Establishes conditions for identification of DDE copula parameters and proves posterior consistency.
Stylized facts can be regarded as constraints for any modeling attempt of price dynamics on a financial market, in that an empirically reasonable model has to reproduce these stylized facts at least qualitatively. The dynamics of market prices is modeled on a macro-level as the result of the dynamic coupling of two dyn…
New Gated Linear Unit (GaLU) improves neural network performance.
problem Theoretical gap between ReLU networks and empirical observations in shallow networks.
method Introduced Gated Linear Unit (GaLU) by decoupling linearity from gating.
result Improved memorization and generalization results for GaLU networks.
dpVAEs improve VAEs by decoupling representation and generation.
problem VAEs struggle with both representation learning and sample generation.
method Introduce decoupled priors (dpVAEs) that separate representation and generation spaces.
result dpVAEs enable regularization without compromising sample generation.
This paper explores the capability of deep neural networks to capture key characteristics of vehicle dynamics, and their ability to perform coupled longitudinal and lateral control of a vehicle. To this extent, two different artificial neural networks are trained to compute vehicle controls corresponding to a reference…