This paper analyzes convergence of DP-SGD with adaptive quantile clipping.
problem Empirical success of adaptive clipping methods lacks theoretical understanding.
method Comprehensive convergence analysis of SGD with quantile clipping (QC-SGD).
result Establishes theoretical guarantees for DP-QC-SGD, revealing relationships between quantile selection, step size, and convergence.
Gradient clipping helps private SGD converge despite potential bias.
problem Gradient clipping in private SGD can bias convergence.
method Theoretical analysis and empirical evaluation of gradient clipping effects.
result Gradient clipping can prevent convergence to stationary points and introduces bias.
Study on gradient clipping in SGD for high-dimensional problems.
problem Understanding and optimizing gradient clipping in high-dimensional machine learning models.
method Theoretical analysis of streaming SGD with gradient clipping in a least squares problem, focusing on large intrinsic dimensionality.
result Developed a deterministic equation to describe the loss evolution in clipped SGD, showing benefits under certain conditions.
A new method improves training speed and robustness for deep learning models.
problem Training deep learning models with heavy-tailed gradient noise.
method Per-sample gradient clipping combined with SGD.
result PS-Clip-SGD achieves optimal convergence rates and high-probability guarantees.
Paper explores whether gradient normalization can replace clipping for SGD in heavy-tailed noise.
problem Ensuring convergence of SGD in heavy-tailed noise.
method Revisits gradient clipping and normalization, proving their sufficiency and effectiveness.
result Gradient normalization alone is sufficient for nonconvex SGD convergence under smoothness assumptions.
Improved Clipped-SGD achieves near-optimal heavy-tailed statistical estimation in streaming settings.
problem High-dimensional heavy-tailed statistical estimation in streaming with memory constraints.
method Stochastic convex optimization with Clipped-SGD, proving near-optimal sub-Gaussian statistical rates.
result Clipped-SGD achieves an error of T r ( Σ ) + T r ( Σ ) ∥ Σ ∥ 2 log ( log ( T ) δ ) T \sqrt{\frac{\mathsf{Tr}(Σ)+\sqrt{\mathsf{Tr}(Σ)\|Σ\|_2}\log(\frac{\log(T)}δ)}{T}} T Tr ( Σ ) + Tr ( Σ ) ∥Σ ∥ 2 l o g ( δ l o g ( T ) ) with probability 1 − δ 1-δ 1 − δ . Large batch training with DP-SGD reduces model performance due to implicit bias.
problem Large batch training with DP-SGD reduces model performance.
method The study analyzes the phenomenon of implicit bias in Noisy-SGD (DP-SGD without clipping) and its theoretical solutions for linear models.
result The implicit bias in large batch training with DP-SGD is amplified by additional noise, similar to SGD.
New method improves convergence of SGD for heavy-tailed noise.
problem Heavy-tailed noise in machine learning applications.
method Normalized SGD (NSGD) to handle heavy-tailed noise.
result Improved sample complexity and high-probability convergence.
Privacy-preserving SGD with heavy-tailed noise achieves differential privacy guarantees.
problem Privacy preservation in noisy SGD with heavy-tailed noise.
method Differential privacy guarantees for SGD with heavy-tailed noise.
result SGD with heavy-tailed perturbations achieves ( 0 , O ( 1 / n ) ) (0, O(1/n)) ( 0 , O ( 1/ n )) -DP. New theory explains why normalization is preferred in SGD under heavy-tailed noise.
problem Understanding why normalization is preferred in stochastic gradient descent (SGD) under heavy-tailed noise.
method Developed a worst-case complexity theory for stochastically preconditioned SGD and its variants.
result Normalization guarantees convergence at optimal rates, while clipping may fail in the worst case.
Privacy preserving machine learning algorithms are crucial for learning models over user data to protect sensitive information. Motivated by this, differentially private stochastic gradient descent (SGD) algorithms for training machine learning models have been proposed. At each step, these algorithms modify the gradie…
Paper proposes DP-SGD and DP-NSGD for differentially private non-convex optimization.
problem Mitigating privacy risks in large model learning.
method Clip or normalize per-sample gradients and add noise for differential privacy.
result Achieved convergence rate of gradient norm for non-convex optimization.
New algorithm robustly optimizes data streams with heavy-tailed or infinite variance samples.
problem Optimizing data streams with heavy-tailed or infinite variance samples.
method Gradient quantile clipping for SGD, leveraging Markov chain connections.
result Algorithm converges to a concentrated distribution with high probability bounds.
AdaDPIGU improves privacy in deep learning by adaptively clipping and pruning gradients.
problem Privacy in deep learning models, especially in high-dimensional settings.
method Importance-based gradient updates, adaptive clipping, differentially private SGD.
result AdaDPIGU achieves high accuracy while maintaining privacy, outperforming non-private models.
A new algorithm reduces communication in deep learning training.
problem Slow convergence in deep neural networks due to exploding gradients.
method Communication-efficient distributed gradient clipping algorithm.
result Algorithm achieves linear speedup and reduced communication rounds.
Private adaptive methods improve on traditional SGD for convex optimization.
problem Differential privacy constraints in gradient optimization.
method Differentially private variants of SGD and AdaGrad with adaptive stepsizes and non-isotropic clipping.
result Private AdaGrad outperforms private SGD in high-dimensional problems.
Improved analysis for clipped gradient methods in nonsmooth convex optimization under heavy-tailed noise.
problem Optimization under heavy-tailed noise in nonsmooth convex problems.
method Refined analysis of Clipped Stochastic Gradient Descent (Clipped SGD) with new rates and improved utilization of Freedman's inequality.
result New rates O ( σ l d m e f f − 1 / 2 p ln 1 − 1 / p ( 1 / δ ) T 1 / p − 1 ) {\cal O}(σ_{\frak l}d_{
m eff}^{-1/2{\frak p}}\ln^{1-1/{\frak p}}(1/δ)T^{1/{\frak p}-1}) O ( σ l d m e f f − 1/2 p ln 1 − 1/ p ( 1/ δ ) T 1/ p − 1 ) and O ( σ l 2 d m e f f − 1 / p ln 2 − 2 / p ( 1 / δ ) T 2 / p − 2 ) {\cal O}(σ_{\frak l}^2d_{
m eff}^{-1/{\frak p}}\ln^{2-2/{\frak p}}(1/δ)T^{2/{\frak p}-2}) O ( σ l 2 d m e f f − 1/ p ln 2 − 2/ p ( 1/ δ ) T 2/ p − 2 ) for nonsmooth convex and strongly convex problems, respectively. New method for differentially private optimization with general Lipschitz conditions.
problem Differentially private optimization under general Lipschitz conditions.
method Generalized Lipschitz condition for per-sample gradients, tuning clip norm based on minimum per-sample Lipschitz constant.
result Efficacy of the recommended clip norm tuning method verified on 8 datasets.
Dp-CLIP preserves privacy in multimodal AI training.
problem Privacy concerns in multimodal AI, especially in vision-language tasks.
method Differentially private adaptation of CLIP model.
result Dp-CLIP retains accuracy while ensuring privacy.
This paper improves analysis of clipping algorithms for non-convex optimization.
problem Improving convergence analysis of clipping algorithms for non-convex optimization.
method Develops a general framework to study clipping algorithms, including momentum methods, and provides convergence analysis in deterministic and stochastic settings.
result Demonstrates that clipping methods can maintain efficiency even in highly non-smooth regions.
New privacy bounds for DP-SGD's last iterate, even with cyclic sampling.
problem Privacy of the last iterate in DP-SGD with cyclic sampling.
method Established new RDP upper bounds for the last iterate under realistic assumptions.
result Privacy bounds for DP-SGD's last iterate with cyclic sampling and clipping, even for nonconvex losses.
This paper improves convergence guarantees for gradient clipping in deep learning.
problem Improving convergence guarantees for gradient clipping in deep learning models.
method Analyzes and provides precise convergence guarantees for arbitrary clipping thresholds.
result Shows tight convergence guarantees for clipped stochastic gradient descent.
New algorithm detects changes in heavy-tailed data streams.
problem Detecting changes in heavy-tailed data streams.
method Clipped Stochastic Gradient Descent (SGD) combined with union bound.
result First algorithm with finite-sample false-positive rate guarantees for heavy-tailed data.
DP-RandP improves privacy-utility tradeoff in DP-SGD by learning priors from random processes.
problem Improving the performance of differentially private stochastic gradient descent (DP-SGD) on private data.
method A three-phase approach that learns priors from images generated by random processes and transfers these priors to private data.
result New state-of-the-art accuracy on CIFAR10, CIFAR100, MedMNIST, and ImageNet for various privacy budgets.
New UCB-type algorithms reduce regret bounds for stochastic bandits with heavy and super heavy noise.
problem Improving regret bounds for stochastic bandits with heavy-tailed noise.
method General convex optimization methods with an inexact oracle, Clipped-SGD-UCB algorithm.
result Achieved an O ( log T K T log T ) O(\log T\sqrt{KT\log T}) O ( log T K T log T ) regret bound for symmetric noise, better than general lower bounds. Improved optimization technique reduces training complexity for non-convex problems.
problem Training non-convex optimization problems with exploding gradients.
method Employed variance reduction technique (SPIDER) with carefully designed learning rate.
result Improved stochastic gradient complexity to O ( ε − 3 ) O(ε^{-3}) O ( ε − 3 ) for ε ε ε -stationary solutions. Improved linear regression with privacy and robustness guarantees.
problem Private and robust linear regression with adversarial corruption.
method Differentially private stochastic gradient descent with full-batch gradient descent and adaptive clipping.
result Near optimal sample complexity for both private and robust linear regression.
Truncated SGD with heavy-tailed noise eliminates sharp local minima.
problem Avoiding sharp local minima in deep learning models.
method Truncated SGD with heavy-tailed gradient noise.
result Truncated SGD can eliminate sharp local minima entirely from its training trajectory.
New bounds for KANs trained with DP-SGD, addressing correlated noise.
problem Risk bounds for Kolmogorov-Arnold Networks trained by DP-SGD with correlated noise.
method Established new optimization and population risk analysis for KANs trained with DP-SGD, addressing correlated noise.
result First optimization and population risk analysis of correlated-noise mechanisms for DP training in non-convex settings, including neural networks.
New method trains deep networks robustly without adaptive methods.
problem Training deep networks with robustness and efficiency.
method Scale invariant architecture + SGD + weight decay + gradient clipping.
result SGD can achieve similar performance to adaptive methods like Adam.
New DP-CD method outperforms DP-SGD in solving composite DP-ERM problems.
problem Privacy-preserving machine learning with differential privacy.
method Differentially Private proximal Coordinate Descent (DP-CD) for composite Empirical Risk Minimization (ERM).
result DP-CD outperforms DP-SGD due to larger step sizes and better gradient exploitation.
Nonlinear SGD achieves high-probability rates in non-convex optimization with heavy-tailed noise.
problem Optimization in non-convex problems with heavy-tailed noise.
method General nonlinear framework for SGD, including symmetrization techniques.
result Achieves O ~ ( t − 1 / 2 ) \widetilde{\mathcal{O}}(t^{-1/2}) O ( t − 1/2 ) rate for heavy-tailed noise. U-Clip clips gradients to reduce bias without discarding them.
problem Reduces bias in stochastic gradient descent.
method Maintains a buffer of clipped gradients to add to next iteration.
result U-Clip updates are unbiased on average.
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.
Differential privacy (DP) is a popular mechanism for training machine learning models with bounded leakage about the presence of specific points in the training data. The cost of differential privacy is a reduction in the model's accuracy. We demonstrate that in the neural networks trained using differentially private …
New method improves fairness in DP learning by preventing excessive gradient suppression.
problem Disparate impact on model predictions for minority groups in DP learning.
method Bounded adaptive clipping to prevent excessive gradient suppression.
result Improves worst-class accuracy by over 10 percentage points compared to existing methods.
Adaptive gradient approaches that automatically adjust the learning rate on a per-feature basis have been very popular for training deep networks. This rich class of algorithms includes Adagrad, RMSprop, Adam, and recent extensions. All these algorithms have adopted diagonal matrix adaptation, due to the prohibitive co…
New RDP guarantees for heavy-tailed SDEs and SGD.
problem Characterizing differential privacy for heavy-tailed noise in learning algorithms.
method Rényi flow computations and fractional Poincaré inequalities.
result First RDP guarantees for heavy-tailed SDEs with weaker dependence on dimension.
CLIP models robustness to spurious features is re-evaluated using a new dataset.
problem Existing robustness tests of CLIP models may not fully reflect their performance on spurious features.
method Crafted a new dataset (CounterAnimal) to reveal CLIP models' reliance on realistic spurious features.
result CLIP models are robust to spurious features learned from their training data, not ImageNet.
CLIP learns joint image-text representations for zero-shot learning.
problem Understanding and improving zero-shot transfer performance in CLIP.
method Formal study of transferrable representation learning and analysis of zero-shot transfer performance.
result Proposes a new CLIP-type approach that outperforms existing methods.
Paper introduces Latent-CLIP for efficient text-image comparison in latent space.
problem Efficiently compare text and images in latent space without costly decoding.
method Trains CLIP model in latent space, uses Latent-CLIP rewards for noise optimization, and guides generation away from harmful content.
result Latent-CLIP matches CLIP performance on text-image classification and harmful content detection.
Enhances CLIP's similarity computation using PMI's linear structure.
problem CLIP's similarity computation misses the optimal linear structure of PMI.
method KME-CLIP, utilizing inner product in a reproducing kernel Hilbert space.
result KME-CLIP approximates PMI with arbitrary accuracy and outperforms CLIP.
AutoClip automatically adjusts gradient clipping for better audio separation.
problem Improving generalization in audio source separation networks.
method Adaptive gradient clipping based on historical gradient norms.
result Improves generalization performance in audio source separation networks.
Unified CLIP space manipulations improve GAN adaptation with a single target image.
problem Overfitting or underfitting in fine-tuning a pre-trained generator with a single target image.
method Two-step training strategy: latent optimization in CLIP space followed by generator fine-tuning with CLIP space consistency loss.
result Our model generates diverse outputs with the target texture and outperforms baseline models.
Existing approaches for training neural networks with user-level differential privacy (e.g., DP Federated Averaging) in federated learning (FL) settings involve bounding the contribution of each user's model update by clipping it to some constant value. However there is no good a priori setting of the clipping norm acr…
Clip21 improves convergence of gradient-clipped methods in DP settings.
problem Gradient clipping introduces bias in distributed training, causing convergence issues.
method Clip21 designs an error feedback mechanism to mitigate bias in gradient-clipped methods.
result Clip21 converges at the same rate as distributed gradient descent, improving from previous $\mathcal{O}\left(\frac{1}{\sqrt{K}}
ight)$ to $\mathcal{O}\left(\frac{1}{K}
ight)$ .
New method for estimating covariance with robustness to outliers.
problem Estimating covariance from noisy data with outliers.
method Cross-fitted clipped covariance estimator with computable Bernstein certificates.
result The method balances certified stochastic error and robust hold-out proxy for clipping bias.
Paper proves convergence for private FL on non-Lipschitz convex objectives using normalization instead of clipping.
problem Lack of convergence results for differentially private federated learning with non-Lipschitz objectives.
method Developed a convergence result for private FL on smooth convex objectives without assuming Lipschitzness, using normalization instead of clipping.
result Normalization-based private FL algorithm converges better than clipping-based counterpart on smooth convex functions.