New bounds enable training of probabilistic models for deep networks.
problem Training scalable latent variable models for deep networks.
method Introducing new variational bounds for specific output layers of neural networks.
result Analytical bounds for certain output layers allow training without re-parameterization or Monte Carlo approximations.
Bounds neural network output distribution to Gaussian for random initialization.
problem Quantifying the distribution of randomly initialized deep neural networks.
method Quantitative Gaussian approximation using quadratic Wasserstein distance.
result Explicit inequalities show how network sizes affect Gaussian behavior.
Two-layer CNNs can overfit well if initialized correctly.
problem Understanding the conditions for benign overfitting in over-parameterized CNNs.
method Extending analysis to fully trainable two-layer CNNs, examining initialization scaling effects.
result Initialization scaling of the output layer is crucial; large scales lead to fixed output behavior, small scales to complex interactions.
New margin measure improves deep learning generalization and robustness.
problem Unclear relationship between output margin and generalization for deep models.
method Introduced 'all-layer margin' for deep neural networks.
result Tighter generalization bounds for neural nets with no exponential depth dependency.
Combines neural networks with variational inference for better uncertainty quantification.
problem Overconfident predictions from traditional neural networks and time-consuming Bayesian optimization.
method VIFO (Variational Inference on the Final-Layer Output) using neural networks to learn mean and variance.
result VIFO provides a good tradeoff in run time and uncertainty quantification, especially for out of distribution data.
Study on deep neural networks using concentration inequalities and optimal stopping.
problem Understanding the performance and structure of stochastic deep neural networks.
method Introduced concentration inequalities for SDNN outputs and an EC classifier. Determined the optimal number of layers via an optimal stopping procedure.
result Optimal number of layers for SDNNs determined via an optimal stopping procedure.
The study shows removing fully connected output layers improves efficiency without sacrificing performance.
problem Large number of parameters in fully connected layers for high-category datasets.
method Examined architectures replacing fully connected output layers with fixed layers and compared performance.
result Fixed classifiers offer no additional benefit over removing the output layer and its parameters.
This paper analyzes and improves the effectiveness of BN techniques in controlling Internal Covariate Shift.
problem The effectiveness of BN techniques in reducing Internal Covariate Shift (ICS) is limited, especially for high-dimensional outputs and noise.
method The paper introduces a measure for ICS using the Earth Mover (EM) distance, derives bounds for this measure, and proposes a unitization algorithm to further bound ICS.
result The unitization algorithm effectively bounds ICS, especially for low-dimensional outputs and small noise, and can be tuned to control ICS.
ReLU networks trained with MILPs match deep learning accuracy.
problem Training deep neural networks efficiently.
method Iterative training with Mixed Integer Linear Programs (MILPs).
result ReLU networks can be trained with MILPs achieving similar accuracy to deep learning methods.
We propose Very Simple Classifier (VSC) a novel method designed to incorporate the concepts of subsampling and locality in the definition of features to be used as the input of a perceptron. The rationale is that locality theoretically guarantees a bound on the generalization error. Each feature in VSC is a max-margin …
One LN layer stabilizes neural network extrapolation.
problem Understanding neural networks' extrapolation behavior.
method Applied NTK theory to infinitely-wide networks.
result Adding one LN layer stabilizes network outputs.
New bounds for neural networks without loss boundedness assumption.
problem Generalization error bounds for two-layer neural networks.
method Wasserstein distance estimates and moment bounds for stochastic gradient method.
result Dimension-free rate of order O(n−1/2) for independent test data. LipKernel adds robustness to CNNs by enforcing Lipschitz bounds.
problem Improving robustness of CNNs in real-time applications.
method Dissipative layers parameterized by LMIs and 2-D Roesser model.
result Orders of magnitude faster run-time compared to state-of-the-art methods.
3-layer NTK models generalize better than 2-layer models, especially with large input dimensions.
problem Understanding the generalization of overparameterized neural networks.
method Analyzing the 3-layer NTK model's test error and comparing it to 2-layer NTK models.
result 3-layer NTK models have a faster descent in test error with respect to the number of neurons in the second hidden layer.
New findings on neural networks with non-negative weights and low training error.
problem Does a low training error imply a small outer norm for two-layer neural networks?
method Covering number argument and fat-shattering dimension analysis.
result For non-negative output weights, low training error guarantees a well-controlled outer norm.
Study efficient neural operator learning using variation spaces.
problem Operator learning using encoder-decoder neural networks.
method Introduce variation space for nonlinear operators, establish approximation bounds.
result Algebraic approximation and learning rates for polynomially decaying input and output encoding errors.
The paper proposes a method to create robust neural networks for automated driving.
problem Creating neural networks that can accurately predict road conditions and distances.
method The method introduces a non-standard loss function with tolerance to account for label variability and allows for deviations from labels.
result The proposed method results in a neural network that can robustly predict road conditions and distances, even with small label variations.
Study on Bayesian deep linear networks with multiple outputs and convolutional layers.
problem Characterize feature learning in finite-width Bayesian deep linear networks.
method Exact and analytical formulas for joint and posterior distributions, using large deviation theory.
result Quantitative description of feature learning in infinite-width regime.
Study clarifies Bayesian generalization error in CBM for 3-layered linear neural networks.
problem Understanding the generalization error in concept bottleneck models.
method Mathematical analysis of Bayesian generalization error and free energy in CBM for 3-layered linear neural networks.
result CBM significantly alters the parameter region and Bayesian generalization error compared to standard models.
In this paper we first present a class of algorithms for training multi-level neural networks with a quadratic cost function one layer at a time starting from the input layer. The algorithm is based on the fact that for any layer to be trained, the effect of a direct connection to an optimized linear output layer can b…
We develop a convex relaxation method for analyzing neural network generalization.
problem Analyzing the generalization of parallel positively homogeneous networks.
method Linking non-convex ERM to a convex optimization problem over prediction functions.
result Achieved generalization bounds with almost linear sample complexity in network width.
We prove that for an L-layer fully-connected linear neural network, if the width of every hidden layer is Ω~(L⋅r⋅dout⋅κ3), where r and κ are the rank and the condition number of the input data, and dout is the output dimension, then gradient descent with Gaussi…
In this paper we introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP …
We introduce a new normalization technique that exhibits the fast convergence properties of batch normalization using a transformation of layer weights instead of layer outputs. The proposed technique keeps the contribution of positive and negative weights to the layer output balanced. We validate our method on a set o…
This paper tightens bounds on the smallest eigenvalue of NTK for deep ReLU networks.
problem Analyzing the smallest eigenvalue of Neural Tangent Kernel for deep ReLU networks.
method Analyzing various quantities of independent interest, including lower bounds on the smallest singular value of hidden feature matrices and upper bounds on the Lipschitz constant of input-output feature maps.
result Tight bounds on the smallest eigenvalue of NTK matrices for deep ReLU nets, both in the limiting case of infinite widths and for finite widths.
The paper proves neural networks with ReLU and softmax can approximate any function.
problem Approximating functions and class labels in neural networks.
method Extended universal approximator theory to neural networks with ReLU and softmax.
result Neural networks with ReLU and softmax can approximate any function and class labels.
Neural ODEs simplified using Chen-Fliess series for Rademacher complexity analysis.
problem Analyzing the complexity of neural ODE models.
method Using Chen-Fliess series to frame neural ODEs as infinite-width nets, where weights are signature of control input and features are Lie derivatives.
result Derived compact expressions for the Rademacher complexity of ODE models.
This research analyzes how input and output layers affect deep neural networks' resistance to adversarial attacks.
problem The vulnerability of deep neural networks to adversarial inputs, especially non-gradient based attacks.
method Analysis of three different fully connected dense network classes with manipulated input and output layers.
result Manipulating input and output layers can significantly enhance a deep neural network's robustness against adversarial attacks.
Paper presents a defense framework against adversarial examples.
problem Vulnerability of deep neural networks to adversarial examples.
method Cross-layer strategic ensemble defense with input and output transformations.
result Strategic ensemble defense achieves high defense success rates and robustness.
This paper examines weight initialization for 1-Lipschitz networks to improve robustness against adversarial attacks.
problem Improving the robustness of deep neural networks against adversarial attacks.
method Examined weight parametrization of AOL and SLL networks, calculated weight variance bounds, and demonstrated weight decay.
result Weight initialization causes deep 1-Lipschitz networks to decay to zero, and weight variance does not affect output variance distribution.
Paper analyzes GCNN sensitivity to probabilistic graph perturbations.
problem Investigating how GCNNs handle probabilistic graph errors.
method Establishes error bounds and linear relationships between GSO perturbations and GCNN outputs.
result GCNNs maintain stability under graph edge perturbations if GSO errors are bounded.
The paper uncovers symmetries in large language models through layer-peeled optimization.
problem Understanding geometric structure in large language model weights and context embeddings.
method Constrained layer-peeled optimization program to analyze symmetries in next-token distributions.
result Symmetries in target next-token distributions are transferred to optimal model weights and context embeddings.
Improved robustness of 1D CNNs for heart arrhythmia classification.
problem Improving the robustness of 1D CNNs for classification tasks.
method Parameterization using Cayley transform and controllability Gramian for Lipschitz-bounded CNNs.
result Improved robustness of trained Lipschitz-bounded 1D CNNs for heart arrhythmia classification.
Proposes BATer for improved adversarial example detection.
problem Detecting adversarial examples in neural networks.
method Introduces a Bayesian adversarial example detector (BATer) using random components in a Bayesian neural network.
result BATer outperforms state-of-the-art detectors in adversarial example detection.
Formalizes explanations as blending input and model output.
problem Creating clear and consistent explanations for model predictions.
method Defines properties of explanation functions and links them to model layers.
result Consistency of activations across layers implies consistency of explanations.
Deep ReLU networks show that 4 layers suffice for unique input recovery.
problem Injectivity capacity of deep ReLU networks.
method Developed a program connecting deep ReLU injectivity to an l-extension of the ℓ0 spherical perceptrons, using random duality theory. result Only 4 layers are needed for unique input recovery, showing expansion saturation effect.
New method improves adversarial training robustness.
problem Lack of tight upper bounds for adversarial training.
method Holistic expansion of the network for upper bound minimization.
result RUB and aRUB methods are more robust than state-of-the-art methods.
We present a framework to derive upper bounds on the number of regions that feed-forward neural networks with ReLU activation functions are affine linear on. It is based on an inductive analysis that keeps track of the number of such regions per dimensionality of their images within the layers. More precisely, the info…
Training large-scale image recognition models is computationally expensive. This raises the question of whether there might be simple ways to improve the test performance of an already trained model without having to re-train or fine-tune it with new data. Here, we show that, surprisingly, this is indeed possible. The …
The paper sets information-theoretic lower bounds for neural networks' parameter recovery and excess risk.
problem Establishing sample complexity lower bounds for neural network parameters and excess risk.
method Using information-theoretic tools, the paper proves lower bounds by constructing a generative network.
result Proves information-theoretic lower bounds for exact parameter recovery and positive excess risk.
Hard problem of learning simple generative models from i.i.d. samples.
problem Learning simple neural network distributions from samples.
method Statistical query model, ODE-based construction of piecewise-linear functions.
result No polynomial-time algorithm can solve this problem even with one-hidden-layer ReLU networks.
We derive upper bounds on the generalization error of learning algorithms based on their \emph{algorithmic transport cost}: the expected Wasserstein distance between the output hypothesis and the output hypothesis conditioned on an input example. The bounds provide a novel approach to study the generalization of learni…
This paper proposes a fast method for estimating input-dependent prediction intervals in Extreme Learning Machines.
problem Estimating reliable prediction intervals for Extreme Learning Machines with heteroscedastic outputs.
method A separate Extreme Learning Machine model estimates input-dependent prediction intervals using a weighted Jackknife method to correct for model uncertainty.
result The proposed method is fast, robust to heteroscedastic outputs, and handles large datasets and insufficient training data.
New L1 regularization controls neural network generalization error and sparsifies input dimensions.
problem Selecting the optimal number of hidden neurons in neural networks.
method Theoretical analysis of L1 regularization in two-layer neural networks. result Appropriate L1 regularization leads to near minimax optimal generalization risk bounds. Interpretability has become an important issue in the machine learning field, along with the success of layered neural networks in various practical tasks. Since a trained layered neural network consists of a complex nonlinear relationship between large number of parameters, we failed to understand how they could achie…
We present a new, unifying approach following some recent developments on the complexity of neural networks with piecewise linear activations. We treat neural network layers with piecewise linear activations as tropical polynomials, which generalize polynomials in the so-called (max,+) or tropical algebra, with pos…
Recently, neural networks in machine learning use rectified linear units (ReLUs) in early processing layers for better performance. Training these structures sometimes results in "dying ReLU units" with near-zero outputs. We first explore this condition via simulation using the CIFAR-10 dataset and variants of two popu…
New method improves BLL models for complex datasets.
problem Limited expressive capacity of Gaussian priors in BLL models.
method Combines diffusion techniques and implicit priors for variational learning.
result Enhanced predictive accuracy and uncertainty quantification.