Study the effects of data parallelism and sparsity on neural network training.
problem Understanding the effects of data parallelism and sparsity on neural network training.
method Conducted extensive experiments and developed a theoretical analysis.
result Found a general scaling trend between batch size and number of training steps to convergence for the effect of data parallelism, and difficulty of training under sparsity.
Cyclic Data Parallelism reduces memory usage and balances gradient communications.
problem Training large deep learning models requires efficient parallelism to scale.
method Cyclic Data Parallelism shifts micro-batches from simultaneous to sequential execution, balancing memory and gradient communications.
result Cyclic Data Parallelism reduces total memory usage and balances gradient communications.
Adaptive batch size schedules improve language model training efficiency and generalization.
problem Dilemma of choosing batch sizes in large-scale model training.
method General-purpose adaptive batch size schedules compatible with data and model parallelism.
result Adaptive batch size schedules outperform constant batch sizes and heuristic warmup schedules.
Adaptive quantization improves SGD accuracy in data-parallel settings.
problem Fixed gradient quantization schemes lead to suboptimal performance in deep learning.
method Developed adaptive quantization schemes ALQ and AMQ that update compression schemes based on gradient statistics.
result Improved validation accuracy on CIFAR-10 and ImageNet datasets by 2% and 1% respectively.
It is time-consuming and error-prone to implement inference procedures for each new probabilistic model. Probabilistic programming addresses this problem by allowing a user to specify the model and having a compiler automatically generate an inference procedure for it. For this approach to be practical, it is important…
Efficiently trains HDP topic models on large datasets using a sparse data-parallel sampler.
problem Scaling non-parametric topic models to large datasets.
method Data-parallel training with a doubly sparse sampler for HDP topic models.
result Trains HDP topic models on a 8m document, 768m token PubMed corpus in under 4 days.
AgEBO-Tabular combines NAS and hyperparameter tuning for fast, high-performing tabular models.
problem Developing high-performing predictive models for large tabular data sets is challenging.
method Combines aging evolution NAS and asynchronous Bayesian optimization for hyperparameter tuning in data-parallel training.
result Automatically discovered neural network models outperform state-of-the-art AutoML ensembles in inference speed by two orders of magnitude.
A new gradient quantization scheme improves communication efficiency in distributed training.
problem Efficiently compressing gradients for parallel training of large models.
method Proposes a new gradient quantization scheme with theoretical guarantees and empirical performance.
result The new scheme matches and exceeds the performance of existing methods.
Proposes NUQSGD for efficient parallel training of large models.
problem Efficiently compressing gradients for parallel SGD training.
method Nonuniform quantization scheme for improved theoretical and empirical performance.
result NUQSGD outperforms QSGDinf and other compression methods.
This work optimizes deep learning training by combining data and model parallelism.
problem Training large models with multiple GPUs suffers from high communication overhead and statistical efficiency loss.
method Hybrid parallelization combining data and model parallelism.
result Hybrid training provides significant speedup compared to data parallelism alone.
Mesh-TensorFlow enables efficient deep learning on large clusters.
problem Memory constraints and inefficiency in batch-splitting for large models.
method Introduces Mesh-TensorFlow for specifying general tensor computations across a multi-dimensional mesh of processors.
result Trains Transformer models with up to 5 billion parameters on TPU meshes of up to 512 cores.
There is significant recent interest to parallelize deep learning algorithms in order to handle the enormous growth in data and model sizes. While most advances focus on model parallelization and engaging multiple computing agents via using a central parameter server, aspect of data parallelization along with decentral…
This paper optimizes deep learning training by efficiently sharding weight updates across replicas.
problem Redundant weight update computation on all replicas in data-parallel training.
method Automatic sharding of weight updates using static analysis and transformations on the training graph.
result Substantial speedups achieved on large-scale models using Cloud TPUs.
Study shows no degradation in neural network performance with larger batch sizes.
problem Characterizing the effects of increasing batch size on neural network training time.
method Experimentally measured training time for various neural network models and datasets.
result No evidence of degradation in out-of-sample performance with larger batch sizes.
Paper proposes DCT for efficient hybrid parallel training of large recommendation models.
problem Training large recommendation models at scale with efficient communication.
method Dynamic Communication Thresholding (DCT) for both Data Parallelism and Model Parallelism.
result Reduces communication by 100x and 20x during DP and MP, respectively, improving training time by 37%.
New method estimates SW distance using CDFs for scalable data parallelism.
problem Estimating SW distance efficiently for large datasets.
method Estimators based on CDFs of projected measures, avoiding sorting.
result Efficient estimation for large datasets and federated learning.
New algorithms speed up EMD computation by four orders of magnitude.
problem High computational complexity of EMD for discrete probability distributions.
method Data-parallel approximation algorithms for linear time complexity.
result Four orders of magnitude faster than existing methods.
In modern large-scale machine learning applications, the training data are often partitioned and stored on multiple machines. It is customary to employ the "data parallelism" approach, where the aggregated training loss is minimized without moving data across machines. In this paper, we introduce a novel distributed du…
The paper optimizes dynamic scheduling for ring architectures in deep learning training.
problem Optimizing deep learning training times with ring architectures.
method Formulated a non-convex, non-linear, NP-hard integer programming problem and developed a doubling heuristic.
result Dynamic scheduling can significantly reduce job completion times in ring architectures.
We determine the critical batch size for large language models and find it scales with data size, not model size.
problem Determining the optimal batch size for large-scale model training.
method We propose a measure of critical batch size, pre-trained models, and systematic hyper-parameter sweeps.
result The critical batch size scales primarily with data size, not model size.
Introduces a natural parallel translation for navigation data.
problem Navigation data geometric representation and parallelism.
method Introduces a natural parallel translation using Riemannian parallelism.
result The natural parallel translation preserves the Randers norm and has a finite-dimensional holonomy group.
We analyze the Hessian spectra of large models up to 100B parameters.
problem Accurate Hessian spectra of large foundation models are difficult to obtain.
method We use shard-local finite-difference Hessian vector products and stochastic Lanczos quadrature.
result We produce the first large-scale spectral density estimates of foundation models.
In real world industrial applications of topic modeling, the ability to capture gigantic conceptual space by learning an ultra-high dimensional topical representation, i.e., the so-called "big model", is becoming the next desideratum after enthusiasms on "big data", especially for fine-grained downstream tasks such as …
This work discusses AAD for financial model calibration and its parallelization benefits.
problem Calibrating stochastic financial models using Automatic Adjoint Differentiation.
method Demonstrates the use of Automatic Adjoint Differentiation for functions in financial models and its parallelization potential.
result Theoretical and numeric results show that AAD allows perfect SIMD parallelization and is efficient.
New framework trains large SciML models solving PDEs in reasonable time.
problem Training large SciML models solving PDEs is challenging and time-consuming.
method Data parallel distributed deep learning framework with optimized methods.
result Neural PDE solvers can be viably trained for practical applications.
We propose a new algorithm called Parle for parallel training of deep networks that converges 2-4x faster than a data-parallel implementation of SGD, while achieving significantly improved error rates that are nearly state-of-the-art on several benchmarks including CIFAR-10 and CIFAR-100, without introducing any additi…
LSAM optimizes deep learning training with improved efficiency.
problem Inefficiency in distributed large-batch training with Sharpness-Aware Minimization (SAM).
method Integrates SAM's adversarial steps with an asynchronous distributed sampling strategy.
result Higher final accuracy compared to data-parallel SAM.
ICSGLD improves efficiency in posterior sampling for big data.
problem Efficient posterior sampling for large datasets.
method Embarrassingly parallel multiple-chain CSGLD with efficient interactions.
result ICSGLD is more efficient than a single-chain CSGLD.
Two efficient ridge solutions improve BLS for new inputs, enhancing accuracy and speed.
problem Improving BLS for new added inputs in a learning system.
method Proposes recursive and square-root BLS algorithms using inverse and inverse Cholesky factor updates.
result Both proposed ridge solutions improve BLS accuracy and speed, especially with larger lambda.
When building large-scale machine learning (ML) programs, such as big topic models or deep neural nets, one usually assumes such tasks can only be attempted with industrial-sized clusters with thousands of nodes, which are out of reach for most practitioners or academic researchers. We consider this challenge in the co…
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.
A simple approach speeds up probabilistic programming models.
problem Efficiently implementing probabilistic programming in deep learning.
method Embedding probabilistic programming into TensorFlow with a single random variable abstraction.
result Optimal linear speedup from 1 to 256 TPUv2 chips, and 100x speedup on GPUs.
Scale of data and scale of computation infrastructures together enable the current deep learning renaissance. However, training large-scale deep architectures demands both algorithmic improvement and careful system configuration. In this paper, we focus on employing the system approach to speed up large-scale training.…
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…
Examines parallels between human subjects and texts for causal inference.
problem Ambiguity and fallacies in causal inference using textual data.
method Two strategies: shifting from traits to perceptions and from concepts to parts.
result Highlights the importance of clarifying fundamental concepts.
Serverless runtimes boost large-scale optimization efficiency.
problem Efficiently solving large-scale optimization problems.
method Master-worker setup with AWS Lambda, parallel optimization algorithm.
result Relative speedups up to 256 workers and efficiencies above 70% up to 64 workers.
PowerSGD compresses gradients for faster distributed optimization.
problem Communication bottleneck in data-parallel distributed optimization.
method Low-rank gradient compressor based on power iteration.
result Achieves test performance on par with SGD and consistent speedups.
SGP combines PushSum with stochastic gradient updates for robust distributed deep learning.
problem Synchronization issues in distributed deep learning.
method Stochastic Gradient Push (SGP) using PushSum for approximate distributed averaging.
result SGP converges to a stationary point at the same rate as SGD and achieves consensus.
A new method reduces data movement in neural network training.
problem Large data movement during neural network training.
method Streaming batch principal component analysis for low-rank updates.
result Effective training of convolutional neural networks with low overhead.
Bayesian non-linear matrix completion tackles large, sparse data.
problem Predict missing elements in large, sparsely observed matrices.
method Bayesian Gaussian process latent variable models with data-parallel distributed computation.
result Scalable Bayesian non-linear matrix completion outperforms linear methods.
TF-Replicator simplifies distributed ML for researchers.
problem Complex distributed ML code for researchers.
method Abstraction over TensorFlow, data-parallel and model-parallel.
result Strong scalability without distributed systems expertise.
Applying machine learning techniques to the quickly growing data in science and industry requires highly-scalable algorithms. Large datasets are most commonly processed "data parallel" distributed across many nodes. Each node's contribution to the overall gradient is summed using a global allreduce. This allreduce is t…
We propose a new integrated method of exploiting model, batch and domain parallelism for the training of deep neural networks (DNNs) on large distributed-memory computers using minibatch stochastic gradient descent (SGD). Our goal is to find an efficient parallelization strategy for a fixed batch size using P process…
New Feedback Transformer architecture improves model performance by exposing past representations to future.
problem Limitations of Transformers in fully exploiting sequential input.
method Proposes Feedback Transformer exposing all past representations to future.
result Demonstrates improved performance with smaller, shallower models.
Adaptive batch sizes improve local gradient methods in distributed training.
problem Communication bottlenecks in distributed deep learning.
method Adaptive batch size strategies for local gradient methods.
result Adaptive batch sizes reduce minibatch gradient variance and improve training efficiency.
We describe a computationally efficient, stochastic graph-regularization technique that can be utilized for the semi-supervised training of deep neural networks in a parallel or distributed setting. We utilize a technique, first described in [13] for the construction of mini-batches for stochastic gradient descent (SGD…
Parallel score matching accelerates DPM training and improves density estimation.
problem Extended training periods and limited modeling flexibility in DPMs.
method Partitioning the learning task into independent time sub-intervals and modeling the score at each time point separately.
result Significant acceleration of training process and improved density estimation performance.
Researchers parallelize neural kernels for large-scale data, achieving state-of-the-art accuracy.
problem Limited scalability of neural kernels on large datasets.
method Massively parallel computation across many GPUs, combined with a distributed, preconditioned conjugate gradients algorithm.
result Achieved state-of-the-art accuracy of 91.2% on CIFAR-5m dataset using neural kernels.