Large scale distributed neural network training through online distillation
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:SZ32ZXJZrecord.jsonopen to challenge →
read the original abstract
Techniques such as ensembling and distillation promise model quality improvements when paired with almost any base model. However, due to increased test-time cost (for ensembles) and increased complexity of the training pipeline (for distillation), these techniques are challenging to use in industrial settings. In this paper we explore a variant of distillation which is relatively straightforward to use as it does not require a complicated multi-stage setup or many new hyperparameters. Our first claim is that online distillation enables us to use extra parallelism to fit very large datasets about twice as fast. Crucially, we can still speed up training even after we have already reached the point at which additional parallelism provides no benefit for synchronous or asynchronous stochastic gradient descent. Two neural networks trained on disjoint subsets of the data can share knowledge by encouraging each model to agree with the predictions the other model would have made. These predictions can come from a stale version of the other model so they can be safely computed using weights that only rarely get transmitted. Our second claim is that online distillation is a cost-effective way to make the exact predictions of a model dramatically more reproducible. We support our claims using experiments on the Criteo Display Ad Challenge dataset, ImageNet, and the largest to-date dataset used for neural language modeling, containing $6\times 10^{11}$ tokens and based on the Common Crawl repository of web data.
This paper has not been read by Pith yet.
Forward citations
Cited by 12 Pith papers
-
Emerging Properties in Self-Supervised Vision Transformers
Self-supervised ViTs show emergent semantic segmentation and 78.3% k-NN accuracy on ImageNet; DINO reaches 80.1% linear evaluation with ViT-Base.
-
Function-Space ADMM for Decentralized Federated Learning: A Control Theoretic Perspective
FedF-ADMM uses function-space ADMM updates projected via knowledge distillation plus a PI-like stabilization term to deliver faster, more stable convergence and higher accuracy than prior decentralized FL methods unde...
-
Enabling Federated Inference via Unsupervised Consensus Embedding
CE-FI maps heterogeneous model representations to a shared embedding space via unsupervised training on unlabeled data, enabling privacy-preserving federated inference that outperforms solo models on image classificat...
-
LatentBurst: A Fast and Efficient Multi Frame Super-Resolution for Hexadeca-Bayer Pattern CIS images
LatentBurst is a new multi-frame super-resolution network for hexadeca-Bayer CIS images that uses pyramid latent alignment, an efficient UNet, and two-step knowledge distillation to handle motion and run on mobile devices.
-
HypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering
HypEHR is a hyperbolic embedding model for EHR data that uses Lorentzian geometry and hierarchy-aware pretraining to answer clinical questions nearly as well as large language models but with much smaller size.
-
The Ratchet Effect in Silico: How Interaction Drives Cumulative Intelligence in Large Language Models
Populations of 1-4B parameter LLMs using peer verification and shared cultural memory achieve 8.8-18.9 point gains on mathematical reasoning tasks and close much of the gap to 70B+ single models.
-
Vision Transformers Need Registers
Adding register tokens to Vision Transformers eliminates high-norm background artifacts and raises state-of-the-art performance on dense visual prediction tasks.
-
Scaling Laws for Task-Specific LLM Distillation
Empirical scaling laws for task-specific LLM distillation in quantitative finance indicate that chain-of-thought supervision recovers general knowledge lost during iterative pruning while in-domain performance degrade...
-
Optimized Federated Knowledge Distillation with Distributed Neural Architecture Search
FedKDNAS combines client-side neural architecture search with knowledge distillation from aggregated server predictions to improve accuracy and efficiency in heterogeneous federated learning.
-
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
Gemini 2.5 Pro and Flash models are presented as achieving frontier performance in reasoning, coding, and long-context multimodal tasks while spanning a cost-capability Pareto curve.
-
Gemma 3 Technical Report
Gemma 3 introduces multimodal open models with architectural changes for efficient long context, trained via distillation and a new post-training recipe that makes the 4B version competitive with prior 27B models and ...
-
Gemma 2: Improving Open Language Models at a Practical Size
Gemma 2 models achieve leading performance at their sizes by combining established Transformer modifications with knowledge distillation for the 2B and 9B variants.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.