Unified spectral framework for μP under joint width-depth scaling.
problem Challenges in stable feature learning and HP transfer for width-depth scaled models.
method Developed a simple and unified spectral framework for μP under joint width-depth scaling.
result Unified and generalized μP formulation for practical architectures with multi-transformation branches.
This paper explores how neural network width and depth behave as they approach infinity.
problem Understanding the behavior of neural functions as width and depth go to infinity.
method Formal definition of commutativity framework, study of neural covariance kernel, novel proof techniques.
result Taking width and depth to infinity in a deep neural network with skip connections results in the same covariance structure, regardless of the order of taking limits.
Theoretical limits of deep residual networks show consistent covariance structures.
problem Understanding the limits of deep residual networks.
method Analyzing the behavior of deep residual networks with skip connections as width and depth approach infinity.
result Theoretical analysis confirms that the covariance structure remains consistent regardless of the order of width and depth.
Study infinite-depth limits of neural networks with fixed width.
problem Understanding the behavior of neural networks as depth increases with fixed width.
method Analyzing finite-width residual networks with random Gaussian weights, focusing on the infinite-depth limit.
result The pre-activations converge to a zero-drift diffusion process, differing from the infinite-width limit.
Self-attention models benefit equally from width and depth, but beyond a certain point, depth becomes less efficient.
problem Understanding the optimal balance between depth and width in self-attention models.
method Theoretical predictions and empirical ablations on networks of varying depths and widths.
result An optimal width of 30K is recommended for a 1-Trillion parameter network, marking a significant width for self-attention models.
New insights into how depth and width affect in-context learning in deep models.
problem Understanding how various resources impact in-context learning in deep models.
method Analyzed linear regression in a deep linear self-attention model, varying resources like depth, width, context length, and training steps.
result Increasing depth improves in-context learning even at infinite context length, contrary to previous findings.
Residual networks with depthwise hyperparameter scaling transfer optimal hyperparameters across width and depth.
problem The challenge of hyperparameter tuning in deep learning, especially for large models.
method Combining μ μ μ P parameterization with residual networks having a residual branch scale of 1 / e x t d e p t h 1/\sqrt{ ext{depth}} 1/ e x t d e pt h . result Optimal hyperparameters transfer across width and depth in residual networks trained with this parameterization.
We prove the precise scaling, at finite depth and width, for the mean and variance of the neural tangent kernel (NTK) in a randomly initialized ReLU network. The standard deviation is exponential in the ratio of network depth to width. Thus, even in the limit of infinite overparameterization, the NTK is not determinist…
ResNets and DenseNets converge to NTK with depth and width, offering advantages for kernel regression.
problem Understanding convergence of ResNets and DenseNets to Neural Tangent Kernel (NTK).
method Analysis of finite width and depth corrections for NTK of ResNets and DenseNets.
result ResNets and DenseNets can converge to NTK with depth and width, unlike vanilla networks.
New Transformer architecture prevents rank degeneracy in deep attention models.
problem Rank degeneracy in deep attention models.
method Modified Softmax-based attention model with skip connections, centered at identity, and scaled logits.
result Existence of a stable SDE implies well-behaved covariance structure, preventing rank degeneracy.
Study on neural network initialization with shaped infinite depth-and-width networks.
problem Understanding the distribution of random covariance matrices in shaped infinite-depth-and-width networks.
method Introduced the Neural Covariance SDE to model the distribution of the random covariance matrix.
result Identified the precise scaling of the activation function necessary for a non-trivial limit.
Inverse depth scaling found in LLMs due to similar layers averaging error.
problem Understanding how depth affects loss in large language models.
method Analysis of LLMs and toy residual networks.
result Loss scales inversely proportional to depth in LLMs.
New scaling framework for MoE architectures ensures stability and optimal performance at scale.
problem Lack of principled understanding of how hyperparameters should scale in MoE architectures.
method Developed a novel Dynamical Mean Field Theory (DMFT) for three scaling regimes of MoE architectures.
result Derived Maximally Scale-Stable Parameterization (MSSP) for SGD and Adam, providing robust learning rate transfer and monotonic improvement with scale.
This work challenges the Neural Tangent Kernel's role in overparameterized neural networks, especially with large width and depth.
problem The Neural Tangent Kernel's behavior in overparameterized neural networks with large width and depth is unclear.
method Experimental and theoretical analysis of ReLU networks with large width and depth.
result The aggregate norm of hidden neuron deviations does not vanish in infinitely-wide ReLU networks, indicating non-trivial behavior.
Adversarial examples in deep ReLU networks with constant depth.
problem Adversarial perturbations lead to large output changes in deep ReLU networks.
method Analyzes the phenomenon in networks with independent Gaussian parameters and constant depth.
result Adversarial examples arise due to functions being close to linear.
Pruned neural networks' error scales predictably with architecture and task.
problem Understanding the predictability of pruning across different scales and architectures.
method Functionally approximated the error of pruned networks, showing it is predictable in terms of invariant tying width, depth, and pruning level.
result The error of pruned networks follows a scaling law with interpretable coefficients that depend on architecture and task.
This paper improves neural network approximation for analytic functions with adjustable depth and width.
problem Approximating analytic functions using neural networks with depth and width parameters.
method Characterizes approximation rates as a joint function of width (N) and depth (L) for ReLU networks.
result Establishes upper bounds for analytic function approximation rates of O(N^(-CL^τ)) with τ influenced by N and L.
Study how depth affects inference in deep Bayesian neural networks.
problem Understanding how depth impacts inference in overparameterized linear Bayesian neural networks.
method Interpreting finite deep linear Bayesian neural networks as scale mixtures of Gaussian process predictors.
result Advances analytical understanding of how depth affects inference in a simple class of Bayesian neural networks.
New findings on depth vs. width in neural networks, showing depth can improve learnability.
problem Understanding the role of depth in neural networks, especially when width is unbounded.
method Analyzing sample complexity for learnability in norm-controlled depth-2 and depth-3 ReLU networks.
result Depth can improve learnability of functions that are otherwise unlearnable with depth-2 networks.
Derives a family of hyperparameter scaling strategies for neural networks.
problem Optimizing hyperparameters for wide and deep neural networks.
method Introduces a one-parameter family of hyperparameter scaling strategies.
result Reveals proper scaling of depth with width for large-scale models.
Improved neural network depth-width trade-offs via dynamical systems.
problem Expressivity of neural networks in terms of depth and width.
method Connection with dynamical systems, focusing on periodic points and Lipschitz constants.
result Sharper width lower bounds for neural networks, yielding exponential depth-width separations.
New findings show depth is more important than width in neural networks.
problem Understanding the role of width and depth in neural networks.
method Constructed networks with bounded weights and width at most d+2, showing depth plays a more significant role.
result Depth is more important than width in the expressive power of neural networks.
CNNs improve generalization to unseen audio devices with increased width, not depth.
problem CNNs are sensitive to specific audio recording devices in acoustic scene classification.
method Investigated the relationship between over-parameterization and generalization in CNNs for audio classification.
result Increasing width improves generalization to unseen devices without increasing the number of parameters.
Deep forests enhance expressiveness exponentially with depth, not width or tree size.
problem Understanding the role of depth, width, and tree size in deep forest performance.
method Provided upper and lower bounds on deep forest approximation complexity.
result Depth exponentially enhances deep forest expressiveness.
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.
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.
ResNets approximate log-Gaussian at initialization, improving network performance.
problem Understanding the initialization behavior of deep neural networks like ResNets.
method Analyzing ReLU ResNets in the infinite-depth-and-width limit, showing log-Gaussian behavior.
result ResNets at initialization exhibit hypoactivation and interlayer correlations, which are not captured by Gaussian limits.
Bayesian linear networks reveal optimal depth and width trade-offs.
problem Understanding how depth, width, and dataset size affect model quality in linear networks.
method Zero noise Bayesian inference with Gaussian weight priors and mean squared error.
result Optimal predictions at infinite depth and maximized Bayesian model evidence at infinite depth.
This article concerns the expressive power of depth in neural nets with ReLU activations and bounded width. We are particularly interested in the following questions: what is the minimal width w min ( d ) w_{\text{min}}(d) w min ( d ) so that ReLU nets of width w min ( d ) w_{\text{min}}(d) w min ( d ) (and arbitrary depth) can approximate any continuous functio…
Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performan…
This paper studies activation sparsity in large language models, finding key trends and implications.
problem Activation sparsity in large language models (LLMs) can be improved for efficiency and interpretability.
method Proposes PPL- p % p\% p % sparsity, analyzes trends with training data, width-depth ratio, and parameter scale. result ReLU is more efficient for sparsity than SiLU, and deeper architectures can improve sparsity.
Estimates neural network error approximating compact sets.
problem Approximating compact subsets from Banach spaces with neural networks.
method Estimates error rates for neural networks of varying width and depth.
result Depth is crucial for better approximation rates, width alone does not improve.
Complex-valued neural networks can approximate any continuous function with bounded widths and depths.
problem Approximating continuous functions with complex-valued neural networks of bounded widths and depths.
method Analyzing activation functions and proving universality for complex-valued networks.
result Deep narrow complex-valued networks are universal if and only if their activation function is neither holomorphic, nor antiholomorphic, nor R \mathbb{R} R -affine. Study of deep linear neural networks with proportional width and depth.
problem Lack of descriptive power in Gaussian limit of deep linear neural networks.
method Proportional infinite-width infinite-depth limit for deep linear neural networks.
result Characterization of limiting distribution as a nontrivial mixture of Gaussians.
Gradient descent converges to global minima for ResNets with linearly scaled width.
problem Understanding the convergence of deep residual networks with varying network width and dataset size.
method Analyzing the Jacobian of ResNets and applying gradient descent for quadratic loss.
result Gradient descent converges to global minima for ResNets with linearly scaled width and independent of depth.
Unified learning-rate scale for CNNs and ResNets, avoiding depth imbalance.
problem Challenges in choosing an appropriate learning rate for deep networks, especially as depth increases.
method Introduces Arithmetic-Mean μ μ μ P (AM- μ μ μ P), constraining network-wide average pre-activation second moment to a constant scale, combined with residual-aware He fan-in initialization. result Demonstrates a − 3 / 2 -3/2 − 3/2 scaling law for learning rates across depths, enabling zero-shot learning-rate transfer. New findings connect shaped and unshaped neural networks using differential equations.
problem Understanding the behavior of neural networks with different activation scaling methods.
method Deriving differential equation-based asymptotic characterizations for shaped and unshaped neural networks.
result Two types of unshaped networks converge to the same infinite-depth-and-width limit at initialization.
This paper studies the expressive power of graph neural networks falling within the message-passing framework (GNNmp). Two results are presented. First, GNNmp are shown to be Turing universal under sufficient conditions on their depth, width, node attributes, and layer expressiveness. Second, it is discovered that GNNm…
This article concerns the expressive power of depth in deep feed-forward neural nets with ReLU activations. Specifically, we answer the following question: for a fixed d i n ≥ 1 , d_{in}\geq 1, d in ≥ 1 , what is the minimal width w w w so that neural nets with ReLU activations, input dimension d i n d_{in} d in , hidden layer widths at most w , w, w , and …
This paper optimizes ReLU networks for approximating Hölder continuous functions.
problem Optimizing the approximation rate of ReLU networks in terms of width and depth.
method Constructive proof of ReLU networks' approximation power with specific width and depth constraints.
result Optimal approximation rate of ReLU networks with width and depth constraints.
New connection between DNNs and Sharkovsky's Theorem for depth-width trade-offs.
problem Understanding why some functions are hard to represent by shallow ReLU networks.
method Connection to Sharkovsky's Theorem and analysis of dynamical systems.
result Lower bounds for width needed to represent periodic functions as a function of depth.
Extends hyperparameter transfer across model sizes and modules, improving training speed.
problem Training stability and performance of large-scale models with optimal hyperparameters.
method Complete ( d ) ^{(d)} ( d ) Parameterisation, per-module hyperparameter optimisation and transfer. result Hyperparameter transfer holds even in the per-module hyperparameter regime, improving training speed.
Deep networks can approximate various activation functions with modest adjustments.
problem Expressive power of deep neural networks with diverse activation functions.
method Approximation of any activation function in set A by ReLU networks with specific scaling factors.
result Approximation of any activation function in a specific subset of A by ReLU networks with (1,1) scaling factors.
Dropout and RaM become equivalent in large ResNets as depth and width increase.
problem Improving performance in deep learning models.
method Comparing Dropout and Random Gradient Masking in ResNets.
result Dropout and RaM converge to the same large-scale limiting dynamics in ResNets.
Deep ReLU networks can approximate smooth functions nearly optimally.
problem Approximating smooth functions with deep neural networks.
method Using Taylor expansions and deep ReLU network approximations, the paper establishes optimal approximation error bounds.
result Deep ReLU networks of width and depth O ( N ln N ) \mathcal{O}(N\ln N) O ( N ln N ) and O ( L ln L ) \mathcal{O}(L\ln L) O ( L ln L ) can approximate f ∈ C s ( [ 0 , 1 ] d ) f\in C^s([0,1]^d) f ∈ C s ([ 0 , 1 ] d ) with an error O ( ∥ f ∥ C s ( [ 0 , 1 ] d ) N − 2 s / d L − 2 s / d ) \mathcal{O}(\|f\|_{C^s([0,1]^d)}N^{-2s/d}L^{-2s/d}) O ( ∥ f ∥ C s ([ 0 , 1 ] d ) N − 2 s / d L − 2 s / d ) . Deep networks with orthogonal weights show stable fluctuations, improving generalization and training speed.
problem Fluctuations in deep networks with Gaussian weights can impair training, especially in networks with depth comparable to width.
method Analytical and numerical studies of fully-connected networks with orthogonal weight initialization and tanh activations.
result Rectangular networks with orthogonal weights have stable fluctuations independent of network depth, leading to better generalization and training speed.
Study on learning rates in neural networks of varying depth.
problem Dependence of maximal update learning rate on network depth.
method Analysis of random fully connected ReLU networks with mean-field weight initialization.
result Maximal update learning rate scales like L − 3 / 2 L^{-3/2} L − 3/2 with network depth. Bayesian neural networks learn efficiently at infinite width, matching polynomial-width performance.
problem Understanding the inductive bias of infinite-width neural networks.
method Analyzing the reduced entropy and using subsampling techniques.
result The Bayesian mean-field learner generalizes exactly on polynomially-bounded targets.