This work investigates the impact of staleness in distributed ML systems and offers insights into convergence.
problem The effects of staleness on the convergence of distributed machine learning algorithms are inconclusive and challenging to monitor.
method Extensive experiments with various ML models and algorithms under delayed updates.
result The empirical findings reveal the diverse effects of staleness on ML algorithm convergence and match the best-known convergence rate.
GA method reduces gradient staleness in cloud computing.
problem Gradient staleness in asynchronous SGD methods.
method Gap-Aware (GA) method that penalizes stale gradients linearly to the Gap.
result GA outperforms existing methods in final test accuracy.
Asynchronous SGD can speed up training with a trade-off of gradient staleness.
problem Asynchronous SGD suffers from gradient staleness, affecting convergence error.
method Theoretical analysis of the error-runtime trade-off considering random straggling delays.
result A method of gradually varying synchronicity in distributed SGD is proposed and demonstrated.
DANA mitigates gradient staleness in asynchronous distributed SGD with momentum.
problem Gradient staleness in asynchronous distributed SGD with momentum.
method DANA: a novel technique for asynchronous distributed SGD with momentum that computes the gradient on an estimated future position of the model's parameters.
result DANA fully incorporates momentum in asynchronous training with almost no ramifications to final accuracy.
Stochastic gradient MCMC (SG-MCMC) has played an important role in large-scale Bayesian learning, with well-developed theoretical convergence properties. In such applications of SG-MCMC, it is becoming increasingly popular to employ distributed systems, where stochastic gradients are computed based on some outdated par…
Distributed Stochastic Gradient Descent (SGD) when run in a synchronous manner, suffers from delays in waiting for the slowest learners (stragglers). Asynchronous methods can alleviate stragglers, but cause gradient staleness that can adversely affect convergence. In this work we present a novel theoretical characteriz…
HPSGD speeds up DNN training by paralleling data sync with local training.
problem Low cluster utilization in distributed deep neural network training.
method Hierarchical Parallel SGD (HPSGD) with improved model updating for stale gradients.
result Significantly boosts distributed DNN training and reduces stale gradients.
DSSP improves deep learning training speed by dynamically adjusting staleness thresholds.
problem Time-consuming deep learning training on large datasets.
method Dynamic Stale Synchronous Parallel (DSSP) framework that adapts staleness threshold at runtime.
result DSSP converges faster and achieves higher accuracy than other paradigms.
Sparsification improves convergence in asynchronous distributed SGD, even in the presence of staleness.
problem Staleness in asynchronous distributed SGD.
method Applied sparsification to reduce communication overheads in distributed asynchronous settings.
result The ergodic convergence rate of sparsified asynchronous SGD matches that of vanilla SGD, $\mathcal{O} \left( 1/\sqrt{T}
ight)$, even in the presence of staleness.
Paper proposes DSP to accelerate deep learning model training by addressing locking and straggler problems.
problem Training deep neural networks is slow and inefficient due to locking and straggler issues.
method Layer-wise Staleness and DSP algorithm to handle locking and straggler problems.
result DSP achieves significant training speedup with stronger robustness than compared methods.
This paper presents Rudra, a parameter server based distributed computing framework tuned for training large-scale deep neural networks. Using variants of the asynchronous stochastic gradient descent algorithm we study the impact of synchronization protocol, stale gradient updates, minibatch size, learning rates, and n…
Asynchronous distributed stochastic gradient descent methods have trouble converging because of stale gradients. A gradient update sent to a parameter server by a client is stale if the parameters used to calculate that gradient have since been updated on the server. Approaches have been proposed to circumvent this pro…
Proposes DC-S3GD for efficient large-scale decentralized neural network training.
problem Training large-scale decentralized neural networks efficiently.
method Decentralized stale-synchronous version of DC-ASGD with gradient correction.
result Achieves state-of-the-art results in training Convolutional Neural Networks.
Unified proof for scalable personalized federated learning.
problem Personalized federated learning under asynchronous updates.
method Unified proof for asynchronous federated learning with bounded staleness applied to MAML and ME personalization frameworks.
result Unified proof for convergence to first-order stationary point for smooth and non-convex functions.
A new method corrects staleness in online optimization by transporting past gradients.
problem Reducing staleness and variance in online optimization methods.
method Implicit gradient transport (IGT) to correct past gradients at the current iterate.
result IGT reduces variance and bias in updates over time and achieves state-of-the-art results.
DG separates successes and failures by gating updates with advantage and surprisal.
problem Negative learning from surprising data in distributed reinforcement learning.
method DG gates each update with the product of advantage and surprisal, suppressing failures and preserving successes.
result DG outperforms other methods in various challenging reinforcement learning tasks.
New framework allows selective removal of stale data in option calibration.
problem Inability to remove old data from calibrated option pricing models without full retraining.
method Introduces operator-theoretic Gauss-Newton framework for selective forgetting.
result Provides stability guarantees and perturbation bounds for selective data removal.
This work addresses the instability in asynchronous data parallel optimization. It does so by introducing a novel distributed optimizer which is able to efficiently optimize a centralized model under communication constraints. The optimizer achieves this by pushing a normalized sequence of first-order gradients to a pa…
Triangle fees adjust fees based on trade size and price movement, improving price accuracy and revenue.
problem Price staleness and low fee revenue in AMMs.
method Decreasing marginal fees proportional to price movement, creating incentives for price accuracy.
result Triangle fees strictly improve the Pareto frontier of price accuracy versus losses.
Asynchronous federated modeling improves spatial data sharing without centralizing raw data.
problem Privacy and bandwidth constraints in distributed spatial data.
method Asynchronous federated modeling using low-rank Gaussian process approximations with block-wise optimization and adaptive strategies.
result Asynchronous federated modeling achieves synchronous performance and outperforms it in heterogeneous settings.
New asynchronous SGD algorithms achieve optimal performance in distributed learning.
problem Asynchronous training introduces staleness, complicating optimization analysis.
method Developed rigorous framework for asynchronous first order stochastic optimization.
result Asynchronous SGD can achieve optimal time complexity, matching synchronous methods.
The paper optimizes cheap and novel flight itineraries using machine learning.
problem Constructing cheap and novel round trip flight itineraries efficiently.
method Formulated as a trade-off between recall and costs, used supervised learning with location embeddings.
result Achieved AUC=80.48, a significant improvement over simpler methods.
A fast method for estimating radar amplitude density parameters.
problem Accurate estimation of amplitude density function parameters in radar applications.
method Projecting amplitude data onto horizontal and vertical axes, then using MLE for α-stale distribution parameters. result The average of computed MLEs based on two projections is a fast and accurate estimator for amplitude distribution parameters.
New model learning objective improves continuous control tasks.
problem Challenges in solving continuous control tasks using model-based reinforcement learning.
method Derived a novel value-aware model learning objective and identified and addressed stale value estimates issue.
result Value-aware objectives can be successfully deployed in solving continuous control tasks without tuning hyper-parameters.
A new algorithm removes stale observations in dynamic Bayesian optimization.
problem Optimizing functions that change over time, keeping track of the optimum.
method Wasserstein distance-based criterion to quantify relevancy, removing stale observations.
result W-DBO maintains good predictive performance and high sampling frequency.
We make policy optimization algorithms batch size-invariant by decoupling proximal and behavior policies.
problem Some policy optimization algorithms do not have batch size-invariance, leading to inefficiencies.
method We decouple the proximal policy from the behavior policy to achieve batch size-invariance.
result Our approach makes policy optimization algorithms more efficient and allows them to use stale data more effectively.
New TVBO algorithm optimizes time-varying functions with varying sampling frequencies.
problem Optimizing time-varying, expensive, noisy functions with constant frequency assumption.
method Formulated practical recommendations and derived upper regret bound for varying sampling frequencies.
result BOLT algorithm outperforms state-of-the-art TVBO algorithms in experiments.
New insights explain speedup saturation in distributed learning with large batches and delays.
problem Understanding and optimizing speedup in distributed learning with large batches and delays.
method Theoretical analysis of strongly convex, convex, and non-convex settings, considering data sparsity.
result Identification of a data-dependent parameter explaining speedup saturation in both batch size and gradient staleness.
New algorithm reduces communication bandwidth for large-scale deep learning training.
problem Efficiently compressing gradients for ring all-reduce in large-scale clusters.
method Importance weighted pruning based on gradient and parameter size.
result Achieved significant gradient compression ratios (64X and 58.8X) on AlexNet and ResNet50.
Unified framework for online learning in click prediction for search ads.
problem Model staleness leading to accuracy and calibration degradation over time.
method Two paradigms of Batch Online Learning: early stopping and proximal regularization.
result Two OL schemes are closely related and can be traded-off between new and historical data.
Rescaled ASGD optimizes distributed learning under heterogeneous data.
problem Vanilla ASGD biases towards a frequency-weighted average of local objectives.
method Rescale worker stepsizes by their computation times.
result Rescaled ASGD converges to the correct global objective in fixed-computation model.
This paper investigates Shampoo's heuristics and decouples preconditioner updates.
problem Improving Shampoo's heuristics for training neural networks.
method Decomposing preconditioner updates, correcting eigenvalues, and adapting eigenbasis computation frequency.
result Principled techniques to remove Shampoo's heuristics and improve training algorithms.
New formula identifies and quantifies costs for automated market makers.
problem Adverse selection costs faced by liquidity providers in automated market makers.
method Derives a Black-Scholes-like formula for AMMs and identifies loss-versus-rebalancing cost.
result Closed-form expressions for LVR applicable to all automated market makers.
Asynchronous computation and gradient compression have emerged as two key techniques for achieving scalability in distributed optimization for large-scale machine learning. This paper presents a unified analysis framework for distributed gradient methods operating with staled and compressed gradients. Non-asymptotic bo…
Asynchronous distributed machine learning solutions have proven very effective so far, but always assuming perfectly functioning workers. In practice, some of the workers can however exhibit Byzantine behavior, caused by hardware failures, software bugs, corrupt data, or even malicious attacks. We introduce \emph{Karda…
We consider parallel asynchronous Markov Chain Monte Carlo (MCMC) sampling for problems where we can leverage (stochastic) gradients to define continuous dynamics which explore the target distribution. We outline a solution strategy for this setting based on stochastic gradient Hamiltonian Monte Carlo sampling (SGHMC) …
MLtuner automatically tunes settings for training tunables (such as the learning rate, the momentum, the mini-batch size, and the data staleness bound) that have a significant impact on large-scale machine learning (ML) performance. Traditionally, these tunables are set manually, which is unsurprisingly error-prone and…
FedBuff improves federated learning scalability with asynchronous updates.
problem Limited scalability of federated learning with synchronous updates.
method Introduces asynchronous updates (staleness) in federated learning.
result Theoretical analysis shows improved convergence rate with boundedness removed.
Twitter uses embedding models to maintain model performance despite covariate shift.
problem Model staleness due to covariate shift in Twitter data.
method Employ learned features (embedding models) to represent relevant aspects of the data distribution.
result Embedding models reduce staleness and maintain model performance.
This paper analyzes error feedback in compressed federated learning for non-convex optimization problems.
problem Reducing communication cost in federated learning with biased gradient compression.
method Proposes Fed-EF, a compressed federated learning scheme with error feedback, and analyzes its convergence rate and performance under partial client participation.
result Fed-EF can match the convergence rate of full-precision FL under data heterogeneity with a linear speedup and no extra slow-down factor due to stale error compensation.
We consider the sequential Bayesian optimization problem with bandit feedback, adopting a formulation that allows for the reward function to vary with time. We model the reward function using a Gaussian process whose evolution obeys a simple Markov model. We introduce two natural extensions of the classical Gaussian pr…
Zeno++ improves robustness of asynchronous SGD in fully asynchronous settings.
problem Byzantine failures in fully asynchronous SGD.
method Estimates descent of loss after applying candidate gradient.
result Proves convergence for non-convex problems under Byzantine failures.
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.
Selective-Backprop accelerates deep learning by focusing on high-loss examples.
problem Training deep neural networks is computationally expensive.
method Prioritizes examples with high loss for gradient computation, skipping others.
result Accelerates training up to 3.5x faster than standard SGD.
A new correction term improves sample efficiency in deep reinforcement learning.
problem Momentum accumulation in TD learning leads to doubly stale gradients.
method Proposed a correction term to address the issue of doubly stale gradients.
result Improves sample efficiency in policy evaluation.
Robo-advisor learns investor's risk preference through portfolio choices.
problem Learning investors' risk preferences without prior knowledge.
method Reinforcement learning framework with exploration-exploitation algorithm.
result Algorithm's value function converges to optimal over polynomial periods.
A model-free framework extracts risk-neutral densities from short-dated options.
problem Arbitrage and bid-ask spread issues in short-dated options.
method Develops ARIES for filtering static arbitrage and SEDEx for density extraction.
result Robust density extraction across various market conditions and volatility smiles construction.
Gradient sparsification reduces communication in deep learning models.
problem Lack of theoretical justification for gradient sparsification methods.
method Proves sparsifying gradients by magnitude with local error correction provides convergence guarantees for data-parallel SGD.
result Gradient sparsification methods can converge under analytic assumptions for both convex and non-convex smooth objectives.