mlsanityData to 7 Oct 2026

Neural networks with late-phase weights

arXiv 2007.12927 ↗First posted 25 Jul 2020cs.LG cs.CV stat.MLcited by 29

Not on the map. Why?

Abstract (arXiv)

The largely successful method of training neural networks is to learn their weights using some variant of stochastic gradient descent (SGD). Here, we show that the solutions found by SGD can be further improved by ensembling a subset of the weights in late stages of learning. At the end of learning, we obtain back a single model by taking a spatial average in weight space. To avoid incurring increased computational costs, we investigate a family of low-dimensional late-phase weight models which interact multiplicatively with the remaining parameters. Our results show that augmenting standard models with late-phase weights improves generalization in established benchmarks such as CIFAR-10/100, ImageNet and enwik8. These findings are complemented with a theoretical analysis of a noisy quadratic problem which provides a simplified picture of the late phases of neural network learning.

Key facts. Cited by 29 papers in the map (data through 7 Oct 2026). First posted on arXiv 25 Jul 2020.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 6; 2022: 6; 2023: 8; 2024: 6; 2025: 1; 2026: 2.

29 papers.

Sort:

By area:

  1. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,366
  2. 2024Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunitiescited by 247
  3. 2023ZipIt! Merging Models from Different Tasks without Trainingcited by 178
  4. 2021Repulsive Deep Ensembles are Bayesiancited by 103
  5. 2021Learning Neural Network Subspacescited by 82
  6. 2021Reinforcement Learning, Bit by Bitcited by 59
  7. 2021Posterior Meta-Replay for Continual Learningcited by 45
  8. 2021Deep Ensembling with No Overhead for either Training or Testing: The All-Round Blessings of Dynamic Sparsitycited by 44
  9. 2023Tangent Model Composition for Ensembling and Continual Fine-tuningcited by 25
  10. 2023Instant Soup: Cheap Pruning Ensembles in A Single Pass Can Draw Lottery Tickets from Large Modelscited by 20
  11. 2023Sparse Model Soups: A Recipe for Improved Pruning via Model Averagingcited by 20
  12. 2024Mix Data or Merge Models? Optimizing for Diverse Multi-Task Learningcited by 17
17 more
  1. 2021Efficient Self-Ensemble for Semantic Segmentationcited by 16
  2. 2024Training-Free Model Merging for Multi-target Domain Adaptationcited by 15
  3. 2022Prune and Tune Ensembles: Low-Cost Ensemble Learning With Sparse Independent Subnetworkscited by 13
  4. 2024Learning More Generalized Experts by Merging Experts in Mixture-of-Expertscited by 12
  5. 2023Adapt then Unlearn: Exploring Parameter Space Semantics for Unlearning in Generative Adversarial Networkscited by 11
  6. 2022Meta-Learning via Classifier(-free) Diffusion Guidancecited by 9
  7. 2022Understanding the effect of sparsity on neural networks robustnesscited by 8
  8. 2022Low-Loss Subspace Compression for Clean Gains against Multi-Agent Backdoor Attackscited by 4
  9. 2022Disentangling the Predictive Variance of Deep Ensembles through the Neural Tangent Kernelcited by 4
  10. 2023A Survey of Historical Learning: Learning Models with Learning Historycited by 3
  11. 2024DIMAT: Decentralized Iterative Merging-And-Training for Deep Learning Modelscited by 3
  12. 2024State Soup: In-Context Skill Learning, Retrieval and Mixingcited by 3
  13. 2023Do the Frankenstein, or how to achieve better out-of-distribution performance with manifold mixing model soupcited by 1
  14. 2025Revisiting Bayesian Model Averaging in the Era of Foundation Modelscited by 1
  15. 2023Improving Robustness of Deep Convolutional Neural Networks via Multiresolution Learningcited by 0
  16. 2026Neural Network Quantization by Learning Low-Loss Subspacescited by 0
  17. 2026A Distributional Optimisation Perspective on Combining Models in Deep Learningcited by 0