Pythia releases 16 identically trained LLMs with full checkpoints and data tools to study training dynamics, scaling, memorization, and bias in language models.
hub Canonical reference
Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
Canonical reference. 82% of citing Pith papers cite this work as background.
85
Pith papers citing it
Background
82% of classified citations
abstract
Deep learning thrives with large neural networks and large datasets. However, larger networks and larger datasets result in longer training times that impede research and development progress. Distributed synchronous SGD offers a potential solution to this problem by dividing SGD minibatches over a pool of parallel workers. Yet to make this scheme efficient, the per-worker workload must be large, which implies nontrivial growth in the SGD minibatch size. In this paper, we empirically show that on the ImageNet dataset large minibatches cause optimization difficulties, but when these are addressed the trained networks exhibit good generalization. Specifically, we show no loss of accuracy when training with large minibatch sizes up to 8192 images. To achieve this result, we adopt a hyper-parameter-free linear scaling rule for adjusting learning rates as a function of minibatch size and develop a new warmup scheme that overcomes optimization challenges early in training. With these simple techniques, our Caffe2-based system trains ResNet-50 with a minibatch size of 8192 on 256 GPUs in one hour, while matching small minibatch accuracy. Using commodity hardware, our implementation achieves ~90% scaling efficiency when moving from 8 to 256 GPUs. Our findings enable training visual recognition models on internet-scale data with high efficiency.
hub tools
citation-role summary
background 15
method 2
citation-polarity summary
claims ledger
- abstract Deep learning thrives with large neural networks and large datasets. However, larger networks and larger datasets result in longer training times that impede research and development progress. Distributed synchronous SGD offers a potential solution to this problem by dividing SGD minibatches over a pool of parallel workers. Yet to make this scheme efficient, the per-worker workload must be large, which implies nontrivial growth in the SGD minibatch size. In this paper, we empirically show that on the ImageNet dataset large minibatches cause optimization difficulties, but when these are address
co-cited works
representative citing papers
Masked autoencoders with asymmetric encoder-decoder and 75% masking ratio enable scalable self-supervised pre-training of vision transformers, achieving 87.8% ImageNet-1K accuracy with ViT-Huge using only unlabeled data.
Self-supervised ViTs show emergent semantic segmentation and 78.3% k-NN accuracy on ImageNet; DINO reaches 80.1% linear evaluation with ViT-Base.
LOSCAR-SGD combines local updates, sparse model averaging, and communication-computation overlap with a delay-corrected merge rule, providing convergence rates for smooth non-convex objectives under worker heterogeneity.
Ringmaster LMO extends delay-thresholding from ASGD to LMO-based momentum updates, providing convergence guarantees under (L0, L1)-smoothness and time-complexity bounds that recover optimal rates in the Euclidean case.
Inkheart SGD and M4 use bidirectional compression to achieve time complexities in distributed SGD that improve with worker count n and surpass prior lower bounds under a necessary structural assumption.
TENNOR enables efficient private training of wide neural networks in TEEs by recasting sparsification as doubly oblivious LSH retrievals and introducing MP-WTA to cut hash table memory by 50x while preserving accuracy.
A new injective multi-jet framework for RL flow control provides jet-count-independent running cost upper bounds and enables superior coordinated jet strategies, achieving drag suppression beyond symmetric ideals on cylinders and aerodynamic efficiency gains from 53% to 73% on airfoils.
Creates the BGTD benchmark and mmTraffic architecture to enable explainable multimodal interpretation of encrypted network traffic using LLMs.
FAERec fuses collaborative ID embeddings with LLM semantic embeddings using adaptive gating and dual-level alignment to enhance tail-item sequential recommendations.
Simulation study shows cold TLB misses in reverse address translation dominate latency for small collectives in multi-GPU pods, causing up to 1.4x degradation, while larger ones see diminishing returns.
FedBCGD reduces communication in federated learning by a factor of 1/N through block-wise parameter updates with accelerated convergence guarantees.
GNEP trains neuroevolution potentials with analytical gradients and Adam optimizer, cutting fitting time by orders of magnitude for Sb-Te systems while matching DFT accuracy on equation of state and radial distribution functions.
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
Mini-batch SGD optimizes a different objective than full partial likelihood in Cox models, but the resulting mb-MPLE is still consistent with optimal rates for neural nets and asymptotic normality for linear models.
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
A promptable model trained on 1B masks achieves competitive zero-shot segmentation performance across tasks and is released publicly with its dataset.
DiTs achieve SOTA FID of 2.27 on ImageNet 256x256 by scaling transformer-based latent diffusion models, with performance improving consistently as Gflops increase.
OPT releases open decoder-only transformers up to 175B parameters that match GPT-3 performance at one-seventh the carbon cost, along with code and training logs.
VICReg prevents collapse in self-supervised image embeddings via explicit variance, invariance, and covariance regularization and matches state-of-the-art downstream performance.
SimCLR learns visual representations by contrasting augmented views of the same image and reaches 76.5% ImageNet top-1 accuracy with a linear classifier, matching a supervised ResNet-50.
Intra-layer model parallelism in PyTorch enables training of 8.3B-parameter transformers, achieving SOTA perplexity of 10.8 on WikiText103 and 66.5% accuracy on LAMBADA.
Switchable Normalization learns per-layer weights to combine channel, layer, and minibatch normalizers, claiming robustness to batch size and better results than fixed normalizers on ImageNet, COCO, CityScapes, ADE20K, MegaFace, and Kinetics.
To achieve robustness to adaptive adversaries, Bernoulli and reservoir sampling require sample size Ω(log |R| / ε²) instead of the static VC-dimension bound.
citing papers explorer
-
LOSCAR-SGD: Local SGD with Communication-Computation Overlap and Delay-Corrected Sparse Model Averaging
LOSCAR-SGD combines local updates, sparse model averaging, and communication-computation overlap with a delay-corrected merge rule, providing convergence rates for smooth non-convex objectives under worker heterogeneity.
-
Ringmaster LMO: Asynchronous Linear Minimization Oracle Momentum Method
Ringmaster LMO extends delay-thresholding from ASGD to LMO-based momentum updates, providing convergence guarantees under (L0, L1)-smoothness and time-complexity bounds that recover optimal rates in the Euclidean case.
-
FedBCD:Communication-Efficient Accelerated Block Coordinate Gradient Descent for Federated Learning
FedBCGD reduces communication in federated learning by a factor of 1/N through block-wise parameter updates with accelerated convergence guarantees.
-
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
-
A Simple Framework for Contrastive Learning of Visual Representations
SimCLR learns visual representations by contrasting augmented views of the same image and reaches 76.5% ImageNet top-1 accuracy with a linear classifier, matching a supervised ResNet-50.
-
Neural Collapse by Design: Learning Class Prototypes on the Hypersphere
Supervised classification reaches neural collapse by design via normalized prototype losses on the hypersphere, outperforming CE and SCL on ImageNet-1K and other benchmarks with faster convergence and better transfer.
-
Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity
Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and bounded heterogeneity.
-
Hypernetworks for Dynamic Feature Selection
Hyper-DFS uses hypernetworks and Set Transformers to generate on-demand parameters for feature subsets in dynamic selection, outperforming prior methods on tabular data and showing stronger zero-shot generalization.
-
OrScale: Orthogonalised Optimization with Layer-Wise Trust-Ratio Scaling
OrScale adds a Frobenius-norm trust-ratio layer-wise scaler to Muon’s orthogonalized updates, with per-layer calibration for language models, yielding higher CIFAR-10 accuracy and better language-model pre-training loss than Muon+Moonlight and AdamW.
-
CommFuse: Hiding Tail Latency via Communication Decomposition and Fusion for Distributed LLM Training
CommFuse eliminates tail latency in communication-computation overlap for distributed LLM training by decomposing collective operations into P2P communications and fusing them with fine-grained computation scheduling.
-
Adaptive Memory Momentum via a Model-Based Framework for Deep Learning Optimization
Presents a model-based proximal framework for adaptive momentum in first-order optimizers by using a two-plane approximation of the objective to dynamically set the memory coefficient online.
-
Mean Flows for One-step Generative Modeling
MeanFlow uses a derived identity between average and instantaneous velocities to train one-step flow models, achieving FID 3.43 on ImageNet 256x256 with 1-NFE from scratch.
-
On the Convergence Theory of Pipeline Gradient-based Analog In-memory Training
Analog-SGD-AP converges with iteration complexity O(ε^{-2} + ε^{-1}) for multi-layer DNNs on AIMC hardware despite analog weight-update imperfections and asynchronous stale gradients.
-
Learning General Representation of 12-Lead Electrocardiogram with a Joint-Embedding Predictive Architecture
ECG-JEPA applies a joint-embedding predictive architecture with Cross-Pattern Attention to learn semantic representations from unlabeled 12-lead ECG data and reports state-of-the-art results on diagnostic classification, feature extraction, and segmentation.
-
Scaling Laws for Transfer
Effective data transferred from pre-training to fine-tuning is described by a power law in model parameter count and fine-tuning dataset size, acting like a multiplier on the fine-tuning data.
-
Tent: Fully Test-time Adaptation by Entropy Minimization
Test-time entropy minimization adapts models by optimizing for confident predictions, reducing error on corrupted ImageNet-C and enabling source-free domain adaptation.
-
Adaptive Federated Optimization
Proposes federated adaptive optimizers (FedAdagrad, FedAdam, FedYogi) with convergence analysis for non-convex objectives under data heterogeneity and reports empirical gains over FedAvg.
-
Large Batch Optimization for Deep Learning: Training BERT in 76 minutes
LAMB optimizer trains BERT with batch size 32868, reducing training time to 76 minutes on TPUv3 Pod without performance loss.
-
Anytime Training with Schedule-Free Spectral Optimization
SF-NorMuon is a new schedule-free spectral optimizer that closes the gap with tuned AdamW on 125M-772M parameter models across 1-8x Chinchilla horizons while providing stationarity guarantees.
-
Fine-Tuning Without Forgetting via Loss-Adaptive Learning Rates
FINCH is a loss-adaptive learning-rate schedule that reduces forgetting by 93% on average during LLM fine-tuning while matching standard task performance across several benchmarks.
-
Information theoretic underpinning of self-supervised learning by clustering
SSL clustering is derived as KL-divergence optimization where a teacher-distribution constraint normalizes via inverse cluster priors and simplifies to batch centering by Jensen's inequality.
-
Multi-Narrow Transformation as a Single-Model Ensemble: Boundary Conditions, Mechanisms, and Failure Modes
Multi-narrow single-model ensembles outperform wide baselines in low-data image classification by learning diverse features but underperform in data-rich settings where training favors few paths.
-
Trust, but Verify: Peeling Low-Bit Transformer Networks for Training Monitoring
A layer-wise peeling framework creates reference bounds to diagnose under-optimized layers in trained decoder-only transformers, including low-bit and quantized versions.
-
Sampling Parallelism for Fast and Efficient Bayesian Learning
Sampling parallelism distributes Bayesian sample evaluations across GPUs for near-perfect scaling, lower memory use, and faster convergence via per-GPU data augmentations, outperforming pure data parallelism in diversity.
-
Gradient Noise Convolution (GNC): Smoothing Loss Function for Distributed Large-Batch SGD
GNC convolves stochastic gradient noise to smooth sharp minima in large-batch SGD, outperforming isotropic noise for better generalization in distributed deep learning.
-
Accelerated Gradient Descent for Faster Convergence with Minimal Overhead
CT-AGD accelerates first-order optimization in deep learning by using finite-difference curvature estimates and noise-mitigation heuristics, achieving equivalent accuracy with 33% fewer training epochs and overhead comparable to Adam.
- LionMuon: Alternating Spectral and Sign Descent for Efficient Training