New method μP2 improves neural network training by scaling perturbations layerwise.
problem Improving neural network performance as models scale up.
method Layerwise perturbation scaling in the infinite-width limit of neural networks.
result Layerwise perturbation scaling ensures all layers are effectively perturbed in the limit.
New framework explains fast transfer of hyperparameters across model scales.
problem Understanding and optimizing hyperparameters for large-scale models.
method Developed a conceptual framework for HP transfer across scale, showing fast transfer is equivalent to useful transfer for compute-optimal grid search.
result Fast transfer of hyperparameters is equivalent to useful transfer for compute-optimal grid search, offering asymptotic computational advantage.
This paper quantifies hyperparameter transfer and finds embedding layer learning rate is key.
problem Quantifying optimal hyperparameters for large language models across scales.
method Developed three metrics to quantify hyperparameter transfer and investigated the importance of embedding layer learning rate.
result Maximal Update (μP) parameterization offers high-quality learning rate transfer compared to standard parameterization (SP).
New geometric interpretation explains over-parameterized models and adversarial perturbations.
problem Geometric understanding of over-parameterized regression and adversarial perturbations.
method Alternative geometric interpretation of regression in feature space.
result Adversarial perturbations are a natural feature of biased models due to underlying geometry.
mSAM improves generalization by making models flatter.
problem Over-parameterized models can have varying generalization performance.
method mSAM modifies loss function to favor flatter minima through adversarial perturbations.
result mSAM achieves flatter minima than SAM and SGD, improving generalization.
Many policy gradient methods are variants of Actor-Critic (AC), where a value function (critic) is learned to facilitate updating the parameterized policy (actor). The update to the actor involves a log-likelihood update weighted by the action-values, with the addition of entropy regularization for soft variants. In th…
Actor-critic methods solve reinforcement learning problems by updating a parameterized policy known as an actor in a direction that increases an estimate of the expected return known as a critic. However, existing actor-critic methods only use values or gradients of the critic to update the policy parameter. In this pa…
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 μP (AM-μP), 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 scaling law for learning rates across depths, enabling zero-shot learning-rate transfer. Framework for safely updating machine learning models.
problem Continuous updates to machine learning models can lead to unintended consequences.
method Formalizes the problem as computing the largest locally invariant domain (LID), uses tractable primal-dual formulation.
result Matches or exceeds heuristic baselines for avoiding forgetting while providing formal safety guarantees.
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.
Graph Neural Networks (GNNs) achieve an impressive performance on structured graphs by recursively updating the representation vector of each node based on its neighbors, during which parameterized transformation matrices should be learned for the node feature updating. However, existing propagation schemes are far fro…
AdvPCA uses robust optimization to achieve sparse PCA without tuning.
problem Sparse PCA for high-dimensional data with implicit sparsity.
method Adversarial PCA (AdvPCA) using robust optimization.
result AdvPCA achieves effective sparse PCA with a closed-form solution.
We develop a general framework for proving rigorous guarantees on the performance of the EM algorithm and a variant known as gradient EM. Our analysis is divided into two parts: a treatment of these algorithms at the population level (in the limit of infinite data), followed by results that apply to updates based on a …
ODS improves adversarial attacks by maximizing output diversity.
problem Efficiency and effectiveness of adversarial attacks, especially black-box attacks.
method Output Diversified Sampling (ODS) that maximizes diversity in model outputs.
result ODS reduces the number of queries needed for black-box attacks on ImageNet by a factor of two.
We propose a new sample-efficient methodology, called Supervised Policy Update (SPU), for deep reinforcement learning. Starting with data generated by the current policy, SPU formulates and solves a constrained optimization problem in the non-parameterized proximal policy space. Using supervised regression, it then con…
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.
While several feature scoring methods are proposed to explain the output of complex machine learning models, most of them lack formal mathematical definitions. In this study, we propose a novel definition of the feature score using the maximally invariant data perturbation, which is inspired from the idea of adversaria…
A new principle for optimizer selection improves training speed and performance.
problem Finding the best optimizer hyperparameters for faster training.
method Formulate optimizer selection as maximizing the expected drop rate in loss, treating gradients and updates as signals and an optimizer as a causal filter.
result Greedy optimizer selection yields stable and effective momentum rules.
Introduces Causal Energy Minimization to understand Transformer layers.
problem Empirical parameterization of Transformer blocks remains largely unexplored.
method Causal Energy Minimization framework that recasts Transformer layers as optimization steps on conditional energy functions.
result Identifies design space for Transformer layers including weight sharing and energy-based interpretations.
Accelerates optimization in asynchronous systems with sparse updates.
problem Optimizing finite-sum objectives in asynchronous lock-free environments.
method New accelerated SVRG variant with sparse updates.
result Achieves optimal incremental gradient complexity.
Gradient perturbation, widely used for differentially private optimization, injects noise at every iterative update to guarantee differential privacy. Previous work first determines the noise level that can satisfy the privacy requirement and then analyzes the utility of noisy gradient updates as in the non-private cas…
In this paper we investigate the usage of adversarial perturbations for the purpose of privacy from human perception and model (machine) based detection. We employ adversarial perturbations for obfuscating certain variables in raw data while preserving the rest. Current adversarial perturbation methods are used for dat…
Deep learning achieves state-of-the-art results in many tasks in computer vision and natural language processing. However, recent works have shown that deep networks can be vulnerable to adversarial perturbations, which raised a serious robustness issue of deep networks. Adversarial training, typically formulated as a …
New variational method solves submodular maximum coverage problem efficiently.
problem Submodular maximum coverage problem in various applications.
method Variational optimization using Nemhauser divergence, alternating E and M steps.
result Efficient solution to submodular maximum coverage problem.
Gradient-based methods introduce noise that penalizes models sensitive to weight perturbations.
problem Noise in gradient-based optimization methods.
method Analysis of Gradient Descent (GD) and Stochastic Gradient Descent (SGD) updating all parameters simultaneously.
result Noise introduced by simultaneous parameter updates penalizes models sensitive to weight perturbations.
A-FADMM improves FL scalability and privacy via wireless channel perturbations and interference.
problem Challenges in model training due to wireless channel randomness and interference.
method Formulated a novel constrained optimization problem and proposed A-FADMM framework.
result Proves convergence and privacy guarantees for A-FADMM under time-varying channels.
Study improves policy search in continuous control by using heavy-tailed distributions.
problem Challenges in continuous space policy search due to non-convexity and myopic-farsighted incentives.
method Introduced heavy-tailed policy parameterizations and analyzed convergence rates and stability.
result Convergence rate to stationarity depends on policy's tail index and exploration tolerance.
We propose a general framework for increasing local stability of Artificial Neural Nets (ANNs) using Robust Optimization (RO). We achieve this through an alternating minimization-maximization procedure, in which the loss of the network is minimized over perturbed examples that are generated at each parameter update. We…
Under-parameterization hinders deep RL's efficiency.
problem Implicit under-parameterization impairs data-efficiency in deep RL.
method Characterized and mitigated the rank collapse of value network features.
result Controlling rank collapse improves deep RL performance.
Learnable token perturbations boost extrapolation in LLMs.
problem Limited flexibility of current discrete perturbations in large language models.
method Learnable continuous latent vector transformations in embedding space, unbiased estimating equations, stochastic gradient descent optimization.
result Significant gains in out-of-domain settings over state-of-the-art methods.
Expectation Maximization (EM) is among the most popular algorithms for maximum likelihood estimation, but it is generally only guaranteed to find its stationary points of the log-likelihood objective. The goal of this article is to present theoretical and empirical evidence that over-parameterization can help EM avoid …
Muon optimizes training efficiency by improving data retention at large batch sizes.
problem Improving training efficiency and data retention at large batch sizes.
method Introducing Muon, a second-order optimizer, and combining it with muP for efficient hyperparameter transfer.
result Muon outperforms AdamW in retaining data efficiency at large batch sizes, enabling more economical training.
Conformal surface parameterization is useful in graphics, imaging and visualization, with applications to texture mapping, atlas construction, registration, remeshing and so on. With the increasing capability in scanning and storing data, dense 3D surface meshes are common nowadays. While meshes with higher resolution …
Constructing an efficient parameterization of a large, noisy data set of points lying close to a smooth manifold in high dimension remains a fundamental problem. One approach consists in recovering a local parameterization using the local tangent plane. Principal component analysis (PCA) is often the tool of choice, as…
This paper re-examines the problem of parameter estimation in Bayesian networks with missing values and hidden variables from the perspective of recent work in on-line learning [Kivinen & Warmuth, 1994]. We provide a unified framework for parameter estimation that encompasses both on-line learning, where the model is c…
New insights into model robustness for random features and NTK models.
problem Understanding and distinguishing robustness in machine learning models.
method Analyzing empirical risk minimization in random features and NTK models.
result Random features models are not robust under any degree of over-parameterization, even when satisfying the universal law of robustness.
In this paper we describe the space of maximal components of the character variety of surface group representations into PSp(4,R) and Sp(4,R). For every rank 2 real Lie group of Hermitian type, we construct a mapping class group invariant complex structure on the maximal components. For the groups PSp(4,R) and Sp(4,R),…
Gradient EM converges globally for over-parameterized Gaussian mixtures.
problem Recovering ground truth Gaussian mixtures with over-parameterized models.
method Gradient EM with over-parameterization, using Hermite polynomials and tensor decomposition.
result Gradient EM globally converges to ground truth with n=Ω(mlogm) over-parameterization. Proposes a probabilistic digital twin for dynamical systems using sparse Bayesian learning.
problem Creating and updating accurate digital twins for complex dynamical systems.
method Sparse Bayesian machine learning, two approaches: input-output and output-only.
result Identifies correct perturbation terms and associated parameters in dynamical systems.
RSO uses random weight perturbations to train deep networks without gradients.
problem Training deep neural networks efficiently and without gradient information.
method RSO is a gradient-free Markov Chain Monte Carlo approach that updates weights based on mini-batch loss reduction.
result RSO achieves high accuracy (99.1% on MNIST) with significantly fewer updates than traditional methods.
Let S be a closed surface of genus at least 2. For each maximal representation ρ:π1(S)→Sp(4,R) in one of the 2g−3 exceptional connected components, we prove there is a unique conformal structure on the surface in which the corresponding equivariant harmonic map to the symmetric spa…
Improves BBVI for high-dimensional Gaussian approximations by using low-rank approximations.
problem Scalability issues with BBVI for high-dimensional multivariate Gaussian approximations.
method Extends BaM framework to handle full covariance matrices by integrating patch step for low-rank parameterization.
result Shows improved efficiency and scalability on synthetic and real-world high-dimensional inference problems.
We prove generic regularity and Uhlenbeck-type compactification theorems for the moduli spaces of PU(2)-monopoles. Generic regularity is NOT obtained in the usual way (by applying Sard theorem to a smooth parameterized moduli space), since the parameterized moduli space can be a priori singular. We explain why, using t…
Noise in SGD affects overparameterized models, favoring sparse solutions.
problem Understanding and mitigating implicit bias in SGD with parameter-dependent noise.
method Theoretical analysis of a quadratically-parameterized model with label noise and Gaussian noise.
result SGD with label noise recovers sparse ground-truth solutions, while SGD with Gaussian noise overfits dense solutions.
We construct and study a family of double-periodic almost entire solutions of the maximal surface equation. The solutions are parameterized by a submanifold of 3×3-matrices (the so-called generating matrices). We show that the constructed solutions are either space-like or of mixed type with the light-cone type…
A framework for federated adversarial learning with convergence analysis.
problem Unique vulnerabilities to adversarial attacks in federated learning.
method Formulates a general federated adversarial learning framework with inner and outer loops for client-side adversarial training and server-side model aggregation.
result The minimum loss under the proposed algorithm can converge to ε with chosen learning rate and communication rounds.
NGD improves multivariate Gaussian inference by optimizing Fisher information.
problem Efficiently optimizing multivariate Gaussian models.
method Natural Gradient Descent applied to multivariate Gaussian parameters.
result NGD updates are more efficient for symmetric covariance matrices.
Study examines infinite limits of transformer dynamics, identifying key parameterizations.
problem Understanding the training dynamics of transformer models in the feature learning regime.
method Analysis of infinite scaling limits using dynamical mean field theory.
result Identified parameterizations that admit well-defined infinite width and depth limits.