Deep networks without non-linearities are equivalent to shallow ones.
problem Training deep orthogonal linear networks with no non-linearity.
method Riemannian gradient descent and gradient descent on factorization.
result Training deep overparametrized networks is equivalent to shallow ones.
Geodesics found in deep linear networks.
problem Finding shortest paths in deep neural networks.
method Derived ODEs and explicit solutions for geodesics.
result Horizontal straight lines are geodesics in invariant manifold.
Two new criteria help understand the advantage of deep neural networks.
problem Understanding the advantage of deepening neural networks.
method Proposed two new criteria to evaluate the expressivity of functions computable by deep neural networks.
result Increasing layers is more effective than increasing units in improving the expressivity of deep neural networks.
Analyzes minima of deep linear networks with weight decay.
problem Understanding the loss landscape of deep neural networks.
method Analytical solutions for global minima with weight decay and stochastic neurons.
result The origin is a special point with qualitatively different minima in networks with more than 1 hidden layer.
Paper converts deep networks to flat, equivalent kernel machines.
problem Capacity control and uniform convergence in deep learning.
method Push-forward transformation from deep networks to indefinite kernel machines.
result Flat network weights are Lp-norm regularized (0<p<1).
Deep linear networks can closely approximate interpolants without improving risk.
problem Understanding the risk bounds of deep linear networks compared to minimum ℓ2-norm solutions. method Bounding excess risk of interpolating deep linear networks trained using gradient flow.
result Deep linear networks can closely approximate or match minimum ℓ2-norm solutions in terms of risk. The paper examines how deep linear neural networks behave as they become infinitely wide.
problem Understanding the behavior of deep linear neural networks as they approach infinite width.
method Analyzes the infinite-width limit of deep linear neural networks, proving convergence to deterministic models and providing precise laws for random weights.
result The training dynamics of deep linear neural networks converge to those of a deterministic model, and the weights' behavior is precisely described.
FA algorithm provides convergence guarantees for deep linear networks.
problem Training efficiency and convergence of deep neural networks.
method Theoretical analysis of Feedback Alignment (FA) algorithm for deep linear networks.
result Certain initializations lead to implicit anti-regularization, affecting learning effectiveness.
Unified entropy formula for real, complex, and quaternionic DLNs.
problem Deriving a formula for DLNs over different fields.
method Extending Menon and Yu's formula to complex and quaternionic DLNs.
result Unified entropy formula for DLNs over R, C, and H. This paper simplifies deep ReLU networks into local linear models for better interpretability.
problem Limited transparency and interpretability of deep neural networks, especially ReLU networks.
method Local linear representation and equivalent set of local linear models (LLMs).
result Simplified deep ReLU networks for better interpretability and diagnostics.
By using the viewpoint of modern computational algebraic geometry, we explore properties of the optimization landscapes of the deep linear neural network models. After clarifying on the various definitions of "flat" minima, we show that the geometrically flat minima, which are merely artifacts of residual continuous sy…
We study the error landscape of deep linear and nonlinear neural networks with the squared error loss. Minimizing the loss of a deep linear neural network is a nonconvex problem, and despite recent progress, our understanding of this loss surface is still incomplete. For deep linear networks, we present necessary and s…
Deep networks become equivalent to linear models in large data regimes.
problem Understanding the behavior of deep neural networks in large data regimes.
method Information-theoretic analysis of fully-trained neural networks in proportional scaling regime.
result Proves deep Gaussian equivalence principle, showing deep networks can be simplified to linear models.
Deep networks prioritize easier examples over harder ones, leading to faster training.
problem Understanding how deep networks prioritize examples of varying difficulty.
method Investigated the effect of linear vs non-linear learning modes on example difficulty.
result Non-linear dynamics tend to sequentialize the learning of examples of increasing difficulty.
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…
Deep Heaviside networks are limited but can be improved with connections or linear neurons.
problem Limited expressivity of deep Heaviside networks.
method Including skip connections or linear activation neurons improves expressivity.
result Lower and upper bounds for VC dimensions and approximation rates are derived.
Sparse linear models improve neural network debuggability.
problem Improving neural network interpretability and debugging.
method Using sparse linear models over learned deep feature representations.
result The approach leads to more debuggable and accurate neural networks.
We investigate the complexity of deep neural networks (DNN) that represent piecewise linear (PWL) functions. In particular, we study the number of linear regions, i.e. pieces, that a PWL function represented by a DNN can attain, both theoretically and empirically. We present (i) tighter upper and lower bounds for the m…
We study the convergence of gradient descent (GD) and stochastic gradient descent (SGD) for training L-hidden-layer linear residual networks (ResNets). We prove that for training deep residual networks with certain linear transformations at input and output layers, which are fixed throughout training, both GD and SGD…
Regularization leads to balancedness in deep linear networks.
problem Balancedness in deep linear networks.
method Geometric invariant theory and Riemannian geometry of fibers.
result Balancing flows converge to the balanced manifold at a uniform exponential rate.
Paper introduces non-linearity signature to measure deep neural network performance.
problem Difficulty in explaining performance differences among similar DNN architectures.
method Affine Optimal Transport mappings to measure non-linearity.
result Signature provides better understanding of DNN inner workings.
New insights into continual learning for deep models, showing convergence issues but local linear solutions.
problem Challenges in continual learning for homogeneous deep models.
method Sequential projections onto task margin sets, leveraging nonconvex projection theory.
result Local linear convergence under certain conditions for homogeneous deep networks.
We consider deep linear networks with arbitrary convex differentiable loss. We provide a short and elementary proof of the fact that all local minima are global minima if the hidden layers are either 1) at least as wide as the input layer, or 2) at least as wide as the output layer. This result is the strongest possibl…
Study approximates nonlinear functionals using deep ReLU networks.
problem Approximating nonlinear continuous functionals with neural networks.
method Constructs continuous piecewise linear interpolation under simple triangulation, analyzes rates of approximation.
result Established rates of approximation for functional deep ReLU networks.
We study the complexity of functions computable by deep feedforward neural networks with piecewise linear activations in terms of the symmetries and the number of linear regions that they have. Deep networks are able to sequentially map portions of each layer's input-space to the same output. In this way, deep models c…
A new estimator learns sparse linear models with context-dependent coefficients.
problem Sparse linear models lack flexibility compared to deep neural networks for handling feature groups.
method Contextual lasso estimator using a deep neural network with lasso regularization.
result Learned models can be sparser than standard lasso without sacrificing predictive power.
Study compares random and learned features in deep Bayesian linear models.
problem Understanding how feature learning affects generalization in deep learning.
method Comparing deep random feature models to deep networks with trained layers.
result Random feature models can display double-descent behavior, while deep networks do not.
DeConvNet, Guided BackProp, LRP, were invented to better understand deep neural networks. We show that these methods do not produce the theoretically correct explanation for a linear model. Yet they are used on multi-layer networks with millions of parameters. This is a cause for concern since linear models are simple …
Lecture notes on linear neural networks for deep learning optimization and generalization.
problem Understanding optimization and generalization in deep learning models.
method Mathematical tools and dynamical systems theory.
result Potential of mathematical tools to enhance understanding of deep learning.
StarNet trains deep models without gradients using linear equations.
problem Training deep generative models with gradients.
method Solving determined systems of linear equations.
result Least-square bounds for latent codes and model parameters.
Study of deep Stable neural networks with various activation functions.
problem Characterizing the infinitely wide limits of deep Stable neural networks.
method Investigation of large-width properties of deep Stable NNs with a generalized central limit theorem for heavy tails.
result Extension of characterization to a broader class of activation functions, including sub-linear, asymptotically linear, and super-linear functions.
Entropy formula derived for deep linear networks using geometric analysis.
problem Thermodynamic description of learning in deep linear networks.
method Group actions, Riemannian submersion, foliation, Jacobi matrices.
result Entropy formula defined on the balanced manifold of DLN.
Deep convolutional networks provide state of the art classifications and regressions results over many high-dimensional problems. We review their architecture, which scatters data with a cascade of linear filter weights and non-linearities. A mathematical framework is introduced to analyze their properties. Computation…
Despite the widespread practical success of deep learning methods, our theoretical understanding of the dynamics of learning in deep neural networks remains quite sparse. We attempt to bridge the gap between the theory and practice of deep learning by systematically analyzing learning dynamics for the restricted case o…
The monotonic linear interpolation in deep networks often leads to plateaus, revealing biases in optimization.
problem Plateaus in the optimization landscape of deep networks during monotonic linear interpolation.
method Investigated monotonic linear interpolation on deep neural networks, focusing on biases in weights and biases.
result Interpolating weights and biases differently can lead to significant differences in loss and accuracy, revealing biases in optimization.
The study analyzes deep linear networks from random initialization, capturing dynamics and hyperparameter effects.
problem Understanding training dynamics in deep linear networks from random initialization.
method Theoretical analysis of gradient descent dynamics in deep linear networks with random initialization and large data.
result Captures the 'wider is better' effect and hyperparameter transfer effects, contrasting with neural-tangent parameterization.
Theoretical comparison of three invariance approaches in deep linear networks.
problem Understanding invariance in deep linear networks.
method Data augmentation, regularization, and hard-wiring approaches.
result Regularization introduces additional critical points, but they remain saddles except for the global optimum.
Use simplified layerwise linear models to understand neural dynamics.
problem Complex neural network dynamics are hard to grasp.
method Apply simplified layerwise linear models to explain neural phenomena.
result Simplified models explain neural collapse, emergence, etc.
Deep linear networks oscillate beyond the edge of stability in a predictable manner.
problem Understanding oscillations in deep linear networks beyond the edge of stability.
method Theoretical analysis of loss oscillations in deep matrix factorization loss.
result Loss oscillations in deep linear networks follow a period-doubling route to chaos and occur within a small subspace.
Study shows directional convergence for neural networks under spherical symmetry.
problem Learning linear predictors with neural networks under spherically symmetric data.
method Analysis of gradient flow and gradient descent for two-layer and deep linear networks.
result Directional convergence guarantees with exact convergence rate for specific network architectures.
We give a covering number bound for deep learning networks that is independent of the size of the network. The key for the simple analysis is that for linear classifiers, rotating the data doesn't affect the covering number. Thus, we can ignore the rotation part of each layer's linear transformation, and get the coveri…
We analyze the global convergence of gradient descent for deep linear residual networks by proposing a new initialization: zero-asymmetric (ZAS) initialization. It is motivated by avoiding stable manifolds of saddle points. We prove that under the ZAS initialization, for an arbitrary target matrix, gradient descent con…
The selection of initial parameter values for gradient-based optimization of deep neural networks is one of the most impactful hyperparameter choices in deep learning systems, affecting both convergence times and model performance. Yet despite significant empirical and theoretical analysis, relatively little has been p…
This work studies learning dynamics in SSMs, linking them to deep linear networks.
problem Lack of theoretical understanding of SSMs, especially in deep state spaces.
method Analyzes learning dynamics of linear SSMs, focusing on frequency domain, and establishes links to deep linear networks.
result Analytical solutions for SSM learning dynamics under mild assumptions, linking to deep linear networks.
Deep networks realize complex mappings that are often understood by their locally linear behavior at or around points of interest. For example, we use the derivative of the mapping with respect to its inputs for sensitivity analysis, or to explain (obtain coordinate relevance for) a prediction. One key challenge is tha…
Paper studies the expressivity of Convolutional Neural Networks (CNNs).
problem Understanding the expressivity of CNNs and their superiority in deep learning.
method Mathematical analysis of linear regions in one-layer and multi-layer ReLU CNNs.
result Deeper CNNs and CNNs have more expressivity per parameter than fully-connected NNs.
Inspired by the recent advances in deep learning (DL), this work presents a deep neural network aided decoding algorithm for binary linear codes. Based on the concept of deep unfolding, we design a decoding network by unfolding the alternating direction method of multipliers (ADMM)-penalized decoder. In addition, we pr…
A new complexity measure for neural networks improves upon classical methods.
problem Lack of a refined complexity measure for comparing different neural network architectures, especially permutation-invariant ones.
method Introduced an equivalence relation among linear functions and counted them relative to this relation.
result The new complexity measure clearly distinguishes between different models and increases exponentially with depth.