Behavior cloning training instabilities amplified by SGD noise over long horizons.
problem Training instabilities in behavior cloning with deep neural networks.
method Empirical dissection of minibatch SGD updates and their effects on long-horizon rewards.
result Exponential moving average (EMA) of iterates effectively mitigates gradient variance amplification (GVA).
Gradient amplification boosts deep learning model performance without increasing training time.
problem Vanishing gradients in deep neural networks.
method Gradient amplification approach to prevent vanishing gradients and training strategy to enable/disable across epochs.
result Improves performance of deep learning models with reduced training time.
SpecGD mitigates misalignment in phase retrieval models with anisotropic inputs.
problem Misalignment during gradient descent in phase retrieval models with anisotropic inputs.
method Spectral gradient descent modifies gradient updates to preserve directional information and remove spike amplification.
result SpecGD removes spike amplification, leading to stable alignment and accelerated noise contraction.
New sampling scheme improves privacy in DP-SGD without sacrificing utility.
problem Suboptimal privacy amplification due to participation variance in Poisson subsampling.
method Balanced Iteration Subsampling (BIS) with structured randomness.
result BIS achieves stronger privacy amplification than Poisson subsampling and is optimal at both extremes of noise spectrum.
We study the phenomenon of bias amplification in classifiers, wherein a machine learning model learns to predict classes with a greater disparity than the underlying ground truth. We demonstrate that bias amplification can arise via an inductive bias in gradient descent methods that results in the overestimation of the…
Sharp pseudospectral bounds prevent transient amplification in coupled gradient descent.
problem Transient amplification in coupled gradient descent systems.
method Developed a sharp pseudospectral theory for block-triangular Jacobians, proving Kreiss constant bounds and matching minimax lower bounds.
result Obtained a finite-horizon iteration-complexity bound of O(K(J)2log(1/δ)) for stochastic coupled descent. Momentum SGD fails to track nonstationary optima due to drift amplification.
problem Tracking nonstationary optima in stochastic optimization.
method Theoretical analysis of SGD and momentum variants under strong convexity and smoothness.
result Momentum incurs a drift-amplification penalty that diverges as the momentum parameter approaches 1, leading to systematic lag.
New method amplifies privacy in decentralized learning without centralized communication.
problem Privacy amplification in decentralized federated learning.
method Random check-in protocol for DP-SGD in FL.
result Privacy/accuracy trade-offs similar to subsampling/shuffling, but without server-initiated communication.
Structured subsampling improves privacy in deep time series forecasting.
problem Incompatible privacy guarantees with time series forecasting.
method Structured subsampling of sequential data for privacy amplification.
result Structured subsampling enables training with strong privacy guarantees.
We investigate the framework of privacy amplification by iteration, recently proposed by Feldman et al., from an information-theoretic lens. We demonstrate that differential privacy guarantees of iterative mappings can be determined by a direct application of contraction coefficients derived from strong data processing…
New algorithm maintains privacy while improving model performance in selective release.
problem Privacy degradation and slow convergence in DPSGD.
method Differentially Private Selective Release based on Clipped Gradients (DPSR-CG).
result Maintains strict privacy guarantees while achieving exceptional model performance.
SGD converges with perturbed forward-backward passes, explained by geometric amplification.
problem Analyzing convergence of SGD with perturbed forward-backward passes in composite optimization.
method Characterized propagation and amplification of perturbations, derived convergence guarantees for non-convex and PL objectives.
result Perturbations cascade through the computational graph, affecting convergence order under specific conditions.
New framework provides privacy guarantees for practical federated learning.
problem Inadequate privacy guarantees for federated learning due to restrictive assumptions.
method Fed-α-NormEC, integrating multiple local updates, partial client participation, and standard assumptions. result Provably convergent and differentially private federated learning framework.
FLAME improves privacy in federated learning without trusted parties.
problem Ensuring privacy in federated learning without trusted parties.
method FLAME uses the shuffle model of differential privacy to achieve better accuracy and privacy.
result FLAME protocols improve testing accuracy by 60.7% compared to local model FL.
New method improves FL efficiency by shuffling data, balancing privacy and accuracy.
problem Balancing privacy, communication, and accuracy in federated learning.
method Developed communication-efficient schemes for private mean estimation, combining privacy amplification and shuffled data.
result Achieved same privacy, optimization performance with lower communication cost.
New analysis shows SGD with noise doesn't leak more privacy with more iterations.
problem Privacy loss in noisy SGD with more iterations.
method Privacy Amplification by Iteration and Sampled Gaussian Mechanism.
result Privacy loss remains constant after a burn-in period, not increasing with more iterations.
Improves privacy amplification by shuffling for differential privacy.
problem Enhancing privacy guarantees in systems with anonymous data contributions.
method Theoretical and numerical analysis of Rényi differential privacy parameters and privacy amplification by shuffling.
result First asymptotically optimal analysis of Rényi differential privacy parameters for shuffled outputs.
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.
Mitigates gender bias amplification in model predictions.
problem Gender bias amplification in model predictions.
method Posterior regularization to mitigate bias.
result Almost removes gender bias amplification in model predictions.
Privacy is enhanced by synthetic data release even with unlimited data.
problem Improving privacy guarantees for synthetic data release.
method Analyzing a bounded-parameter assumption to show privacy amplification persists with unlimited synthetic records.
result Privacy amplification is possible even with an unbounded number of synthetic records.
AIS corrects rollout-training mismatch in quantized RL, improving speed and stability.
problem Rollout-training mismatch in quantized RL causes bias and training collapse.
method Adaptive Importance Sampling (AIS) adjusts gradient correction per batch.
result AIS matches BF16 baseline on most tasks while improving speed.
New method shows hidden state can significantly improve differential privacy in SGD.
problem Differential privacy in SGD with hidden state.
method Proves converging privacy bounds for hidden state SGD, using privacy amplification techniques.
result Privacy bound converges exponentially fast and is smaller than composition bounds.
Paper proves privacy guarantees for shuffled and online PNSGD, reducing noise over time.
problem Privacy amplification in shuffled and online PNSGD settings.
method Iterative analysis of PNSGD with hidden updates, proving privacy guarantees for shuffled and online settings.
result Privacy guarantees for shuffled and online PNSGD with reduced noise over time.
Study shows gradient variance increases during deep learning training, contrary to common belief.
problem Understanding and minimizing gradient variance in deep learning models.
method Gradient Clustering method using stratified sampling to minimize gradient variance.
result Gradient variance increases during training, and smaller learning rates coincide with higher variance.
Paper improves differential privacy analysis for machine learning.
problem Quantifying privacy leakage in noisy gradient descent.
method Shifted interpolation in f-differential privacy. result First exact privacy analysis for strongly convex optimization.
Many commonly used learning algorithms work by iteratively updating an intermediate solution using one or a few data points in each iteration. Analysis of differential privacy for such algorithms often involves ensuring privacy of each step and then reasoning about the cumulative privacy cost of the algorithm. This is …
In this article we quantify the bullwhip effect (the variance amplification in replenishment orders) when demands and lead times are predicted in a simple two-stage supply chain with one supplier and one retailer. In recent research the impact of stochastic order lead time on the bullwhip effect is investigated, but th…
This work proposes using zero-variance control variates to reduce variance in pathwise gradient estimators for variational inference.
problem Pathwise gradient estimators in variational inference have high variance, leading to inefficient optimization.
method Apply zero-variance control variates to pathwise gradient estimators.
result Zero-variance control variates can significantly reduce the variance of pathwise gradient estimators without requiring complex assumptions.
VarGrad reduces variance in ELBO gradient estimation for variational inference.
problem Improving the variance of gradient estimators in variational inference.
method VarGrad uses a new log-variance loss to estimate the ELBO gradient, achieving lower variance than the score function method.
result VarGrad offers a lower variance gradient estimator compared to other methods.
Improved shuffling technique amplifies privacy guarantees for anonymous data contributions.
problem Enhancing privacy in systems where data is contributed anonymously.
method Developed a new approach to random shuffling that amplifies differential privacy guarantees.
result Achieved asymptotically optimal privacy amplification with nearly optimal dependence in ε.
Novel privacy model for decentralized data analysis.
problem Achieving good privacy-utility trade-off in federated learning.
method Introducing network Differential Privacy (network DP) for decentralized algorithms.
result Privacy-utility trade-offs of network DP algorithms significantly improve upon LDP and trusted curator model.
Latent-state environments with long horizons, such as those faced by recommender systems, pose significant challenges for reinforcement learning (RL). In this work, we identify and analyze several key hurdles for RL in such environments, including belief state error and small action advantage. We develop a general prin…
Improves gradient estimation for discrete distributions with variance reduction techniques.
problem Excessive variance in gradient estimation for discrete distributions.
method Stein operators for discrete distributions and control variates.
result Substantially lower variance in gradient estimation.
New method reduces inference variance for faster optimization.
problem High variance in black-box variational inference.
method Joint control variate addressing both data subsampling and Monte Carlo noise.
result Significantly reduced gradient variance, leading to faster optimization.
Large batch sizes reduce gradient variance in DP-SGD, improving privacy.
problem Understanding why large batch sizes work in DP-SGD.
method Decomposed total gradient variance into subsampling and noise-induced variances, proving batch size independence in the limit.
result Large batch sizes reduce effective total gradient variance, improving privacy in DP-SGD.
Gradient-based Monte Carlo sampling algorithms, like Langevin dynamics and Hamiltonian Monte Carlo, are important methods for Bayesian inference. In large-scale settings, full-gradients are not affordable and thus stochastic gradients evaluated on mini-batches are used as a replacement. In order to reduce the high vari…
The paper studies stochastic gradient descent with infinite variance gradients.
problem Theoretical properties of SGD with infinite variance gradients.
method Establish asymptotic behavior of SGD with infinite variance gradients.
result Asymptotic distribution of SGD is characterized as a stationary distribution of an Ornstein-Uhlenbeck process driven by a stable Lévy process.
A fundamental result in differential privacy states that the privacy guarantees of a mechanism are preserved by any post-processing of its output. In this paper we investigate under what conditions stochastic post-processing can amplify the privacy of a mechanism. By interpreting post-processing as the application of a…
New bounds show BBVI's gradient variance matches SGD conditions, improving parameterization efficiency.
problem Understanding and improving the convergence of black-box variational inference (BBVI).
method Showed BBVI satisfies matching gradient variance bounds corresponding to the ABC condition for smooth and quadratically-growing log-likelihoods.
result Proven BBVI's gradient variance matches SGD conditions, with superior dimensional dependence for mean-field parameterization.
Many real world learning tasks involve complex or hard-to-specify objectives, and using an easier-to-specify proxy can lead to poor performance or misaligned behavior. One solution is to have humans provide a training signal by demonstrating or judging performance, but this approach fails if the task is too complicated…
A new gradient estimator reduces variance near boundaries for binary latent variables.
problem Explosive gradient variance near boundaries in binary latent variable models.
method Introduces a new gradient estimator (bitflip-1) and an aggregated estimator (UGC) that uses either bitflip-1 or DisARM for each coordinate.
result UGC has uniformly lower variance than DisARM and achieves optimal optimization objectives.
Gradient descent amplifies random features in neural networks to useful ones.
problem Generalization in neural networks trained on corrupted data.
method Characterization of feature-learning process in two-layer ReLU networks trained by gradient descent.
result Gradient descent amplifies random features to useful ones, achieving near optimal generalization error.
We consider the core reinforcement-learning problem of on-policy value function approximation from a batch of trajectory data, and focus on various issues of Temporal Difference (TD) learning and Monte Carlo (MC) policy evaluation. The two methods are known to achieve complementary bias-variance trade-off properties, w…
New method for privacy amplification without sampling for matrix factorization.
problem Privacy amplification for differentially private model training with matrix factorization.
method Sampling-free bounds based on Rényi divergence and conditional composition.
result Stronger privacy guarantees for small ε, applicable to various matrices.
Unified framework for subsampling mechanisms with tighter privacy guarantees.
problem Improving privacy in machine learning models through subsampling.
method Conditional optimal transport for deriving mechanism-specific subsampling guarantees.
result Tighter privacy bounds for subsampled mechanisms compared to traditional methods.
Improved SGD bounds for machine learning models with Markovian noise.
problem Uniform high-probability bounds for SGD under PL condition with Markovian noise.
method Combining Poisson equation for Markovian noise and probabilistic induction for almost-sure bounds.
result Matching 1/k decay rate for expected suboptimality. This paper tackles variance issues in GNN training by proposing a method to reduce both embedding and gradient variances.
problem High variance in estimating stochastic gradients in GNN training, especially in large graphs.
method The paper proposes a decoupled variance reduction strategy that employs approximate gradient information to adaptively sample nodes with minimal variance.
result The proposed method achieves faster convergence and better generalization compared to existing sampling methods.
The paper develops a theory explaining how machine learning models can amplify biases.
problem Understanding and mitigating bias in machine learning models.
method Analytical theory of ridge regression with and without random projections.
result Observations and predictions align with empirical data on machine learning bias.