Research
On-device research index

arXiv research

A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.

168,932 papers · 148 categories

Trend · papers per month

3977116154 · Jun 202019922001200920172026
48 results for per-layer update strength

LoRA fine-tuning creates intruder dimensions that can cause forgetting, and a new law predicts when this happens.

problem Predicting when LoRA fine-tuning creates intruder dimensions that can cause catastrophic forgetting.
method Derived a per-layer critical update strength ss^\ast and an exact secular-equation characterization of the updated spectrum.
result The law localizes the empirical threshold within a factor of two on 82% of layers and separates intruder-bearing from intruder-free layers at deployment.

Improved differentially private deep learning with group-wise clipping techniques.

problem Efficiency and privacy trade-offs in deep learning models.
method Group-wise clipping techniques (per-layer and per-device) to reduce compute time and memory overhead.
result Private learning with group-wise clipping achieves similar or better performance than non-private learning with less wall time.

A new DP method for deep learning with faster convergence and better privacy.

problem Challenges in differentially private training of deep neural networks.
method Method of auxiliary coordinates with perturbed Taylor expansion for privacy.
result Empirically shows decent trained model quality with modest privacy budget.

Selective state-adaptive regularization improves offline RL performance.

problem Extrapolation errors and value overestimation in static dataset RL.
method State-adaptive regularization coefficients trust Bellman-driven results selectively.
result Significant improvement in performance on D4RL benchmark.

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.

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 μμP (AM-μμP), 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 scaling law for learning rates across depths, enabling zero-shot learning-rate transfer.

SAEs struggle with curved activation manifolds, revealing layer-dependent scaling laws.

problem Sparse autoencoders' reconstruction error varies across layers, not fitting existing scaling laws.
method Cross-layer study of 844 SAE checkpoints, fitting and regressing on manifold geometry.
result Manifold geometry predicts layer-dependent width exponents in SAEs, with transferable coefficients.

Recent pruning methods at initialization fall short of random pruning's accuracy.

problem Improving neural network accuracy through pruning at initialization.
method Various pruning methods (SNIP, GraSP, SynFlow, magnitude pruning) are evaluated; per-layer pruning decisions are proposed.
result Randomly shuffling or sampling initial weights preserves or improves accuracy, suggesting challenges with pruning heuristics.

Adaptive sampling method optimizes DNN compression for resource-constrained platforms.

problem Efficiently compressing DNNs for resource-constrained platforms with high accuracy.
method Adaptive sampling using genetic algorithm-inspired operations to optimize hyperparameters.
result Adaptive sampling outperforms rule-based and reinforcement learning methods in compression rate and accuracy.

A new tensor network method for image classification reduces computation cost.

problem Efficiently classifying images in high-dimensional spaces.
method Proposes a multi-layered tensor network (MLTN) that performs one MPS operation per layer, reducing computation cost.
result Reduces computation cost without degrading performance.

RELTA-SGLD stabilizes nonconvex SGLD updates with a lighter taming scheme.

problem Stabilizing superlinear stochastic-gradient updates in nonconvex optimization.
method Threshold-based taming with relative-growth principle for stability.
result Polynomial moment stability and first-order stationary accuracy in nonconvex SGLD.

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.

FedCONST adapts update magnitudes to enhance feature generalization in FL.

problem Heterogeneous client data in FL leads to overfitting and distorted transferable features.
method FedCONST uses linear convex constraints to stabilize training and preserve generalization.
result FedCONST enhances feature transferability and robustness, achieving state-of-the-art performance.

Examines learned optimizers in deep learning, revealing their dynamics and potential improvements.

problem Stability and generalization issues with learned optimizers.
method Analyzes optimization trajectories using network architecture symmetries and parameter update distributions, contrasts with manual designs.
result Identifies key insights for improving learned optimizers by leveraging strengths from both learned and manual designs.

A new method automatically and dynamically sets learning rates in deep learning.

problem Determining the appropriate learning rate in deep learning tasks is challenging and often subjective.
method Local Quadratic Approximation (LQA) to automatically and dynamically set learning rates.
result The proposed method leads to nearly optimal learning rates in a computationally efficient way.

Two algorithms learn Gaussian graphical models from Glauber dynamics trajectories, achieving optimal performance.

problem Learning Gaussian graphical models from a single trajectory of a dependent stochastic process.
method Two algorithms based on dueling-neighborhood search and local statistics built from the update sequence of Glauber dynamics.
result Achieve κ2κ^{-2} dependence of the information-theoretic lower bounds, mixing-free and signal-optimal.

One of the main difficulties in analyzing neural networks is the non-convexity of the loss function which may have many bad local minima. In this paper, we study the landscape of neural networks for binary classification tasks. Under mild assumptions, we prove that after adding one special neuron with a skip connection…

2018-05-22abs ↗pdf ↗

Jeffrey guidance extends diffusion-model control to more complex applications.

problem Controlling diffusion models beyond simple cases like conditional sampling.
method Leveraging Jeffrey's rule of conditioning to update marginal distributions towards a target distribution.
result Significant reductions in FID on CIFAR-10 and FFHQ with Inception embeddings as the target.

A new learning strategy using two GP layers for inhomogeneous data.

problem Addressing inhomogeneous empirical correlation structures in data.
method Modeling the function as a sample function of a non-stationary Gaussian Process (GP) nested within multiple stationary GPs, with hyperparameters dependent on the outer GP.
result The approach is sufficient with two GP layers, and the model can be implemented using MCMC.

We propose a voted dual averaging method for online classification problems with explicit regularization. This method employs the update rule of the regularized dual averaging (RDA) method, but only on the subsequence of training examples where a classification error is made. We derive a bound on the number of mistakes…

2013-10-17abs ↗pdf ↗

Consistent estimator derived for confounding strength in observational data.

problem Estimating confounding strength in observational data is challenging due to unobserved confounders.
method Derived and adapted a consistent estimator using tools from random matrix theory.
result The original estimator is not consistent, but an adapted one is.

E2^2M optimizes tensor density estimation by relaxing αα-divergence to KL-divergence.

problem Analytical challenges in traditional αα-divergence optimization for tensor-based density estimation.
method E2^2M algorithm: relaxes optimization to KL-divergence, then applies tensor many-body approximation.
result Flexible modeling of various low-rank structures and their mixtures.

SONIA optimizes machine learning problems with a novel algorithm.

problem Empirical risk minimization in machine learning.
method Symmetric Blockwise Truncated Optimization (SONIA) algorithm combining second-order and steepest descent steps.
result SONIA converges to stationary points in both convex and nonconvex cases.

GRPO optimizes LLMs with verifiable rewards, amplifying policy success.

problem Improving LLMs' reasoning under verifiable binary rewards.
method Introduces GRPO, analyzes variants of reward normalization and regularization.
result GRPO amplifies policy success, converging to a fixed point exceeding the reference.

This work optimizes RL algorithms using entropy regularisation for continuous-time LQ problems.

problem Designing RL algorithms to balance exploration and exploitation in noisy environments.
method Entropy regularisation in two formulations: exploratory control and proximal policy update.
result Regret of O(N)\mathcal{O}(\sqrt{N}) for both learning algorithms over NN episodes.

Understanding tie strength in social networks, and the factors that influence it, have received much attention in a myriad of disciplines for decades. Several models incorporating indicators of tie strength have been proposed and used to quantify relationships in social networks, and a standard set of structural networ…

2017-10-11abs ↗pdf ↗

Novel method uses information theory to measure causal influences during transient neural events.

problem Characterizing network interactions during transient neural events.
method Structural Causal Models, Information Theory, Transfer Entropy, Dynamic Causal Strength, Relative Dynamic Causal Strength.
result Introduced a novel measure, relative Dynamic Causal Strength, with theoretical and empirical support.

Unified Latent Dynamics unifies model-free and model-based reinforcement learning.

problem Combining the efficiency of model-free methods with the representational strengths of model-based approaches.
method Embedding state-action pairs into a latent space where the true value function is approximately linear, using synchronized updates of encoder, value, and policy networks.
result ULD achieves cross-domain competence with minimal tuning and a fraction of the parameter footprint.

This paper calculates interaction strength for translation surfaces with multiple singularities.

problem Computing the interaction strength of translation surfaces with multiple singularities is challenging.
method The authors study interaction strength of specific families of translation surfaces, including regular polygons and Bouw-Möller surfaces.
result The paper provides exact computations of KVol on translation surfaces with multiple singularities.

For any positive integer kk, there exist neural networks with Θ(k3)Θ(k^3) layers, Θ(1)Θ(1) nodes per layer, and Θ(1)Θ(1) distinct parameters which can not be approximated by networks with O(k)\mathcal{O}(k) layers unless they are exponentially large --- they must possess Ω(2k)Ω(2^k) nodes. This result is proved here for a class o…

2016-02-14abs ↗pdf ↗

NVA combines variational posteriors, annealing, and natural-gradient learning for multimodal optimization.

problem Finding multiple global and local modes in nonconvex objectives.
method NVA integrates variational posteriors, annealing, and natural-gradient learning.
result NVA outperforms gradient descent and evolution strategies on simulations and real-world problems.