Scaling Exponents Across Parameterizations and Optimizers

• URL: http://arxiv.org/abs/2407.05872v2
• Date: Tue, 16 Jul 2024 17:40:09 GMT
• Title: Scaling Exponents Across Parameterizations and Optimizers
• Authors: Katie Everett, Lechao Xiao, Mitchell Wortsman, Alexander A. Alemi, Roman Novak, Peter J. Liu, Izzeddin Gur, Jascha Sohl-Dickstein, Leslie Pack Kaelbling, Jaehoon Lee, Jeffrey Pennington,
• Abstract summary: We propose a new perspective on parameterization by investigating a key assumption in prior work. Our empirical investigation includes tens of thousands of models trained with all combinations of threes. We find that the best learning rate scaling prescription would often have been excluded by the assumptions in prior work.
• Score: 94.54718325264218
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Robust and effective scaling of models from small to large width typically requires the precise adjustment of many algorithmic and architectural details, such as parameterization and optimizer choices. In this work, we propose a new perspective on parameterization by investigating a key assumption in prior work about the alignment between parameters and data and derive new theoretical results under weaker assumptions and a broader set of optimizers. Our extensive empirical investigation includes tens of thousands of models trained with all combinations of three optimizers, four parameterizations, several alignment assumptions, more than a dozen learning rates, and fourteen model sizes up to 26.8B parameters. We find that the best learning rate scaling prescription would often have been excluded by the assumptions in prior work. Our results show that all parameterizations, not just maximal update parameterization (muP), can achieve hyperparameter transfer; moreover, our novel per-layer learning rate prescription for standard parameterization outperforms muP. Finally, we demonstrate that an overlooked aspect of parameterization, the epsilon parameter in Adam, must be scaled correctly to avoid gradient underflow and propose Adam-atan2, a new numerically stable, scale-invariant version of Adam that eliminates the epsilon hyperparameter entirely.

Related papers

• ETHER: Efficient Finetuning of Large-Scale Models with Hyperplane Reflections [59.839926875976225]
  We propose the ETHER transformation family, which performs Efficient fineTuning via HypErplane Reflections. In particular, we introduce ETHER and its relaxation ETHER+, which match or outperform existing PEFT methods with significantly fewer parameters.
  arXiv  Detail & Related papers  (2024-05-30T17:26:02Z)
• A Unified Gaussian Process for Branching and Nested Hyperparameter Optimization [19.351804144005744]
  In deep learning, tuning parameters with conditional dependence are common in practice. New GP model accounts for the dependent structure among input variables through a new kernel function. High prediction accuracy and better optimization efficiency are observed in a series of synthetic simulations and real data applications of neural networks.
  arXiv  Detail & Related papers  (2024-01-19T21:11:32Z)
• Should We Learn Most Likely Functions or Parameters? [51.133793272222874]
  We investigate the benefits and drawbacks of directly estimating the most likely function implied by the model and the data. We find that function-space MAP estimation can lead to flatter minima, better generalization, and improved to overfitting.
  arXiv  Detail & Related papers  (2023-11-27T16:39:55Z)
• Parameter-Efficient Fine-Tuning without Introducing New Latency [7.631596468553607]
  We introduce a novel adapter technique that directly applies the adapter to pre-trained parameters instead of the hidden representation. Our proposed method attains a new state-of-the-art outcome in terms of both performance and storage efficiency, storing only 0.03% parameters of full fine-tuning.
  arXiv  Detail & Related papers  (2023-05-26T08:44:42Z)
• Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning [91.5113227694443]
  We propose a novel visual. sensuous-aware fine-Tuning (SPT) scheme. SPT allocates trainable parameters to task-specific important positions. Experiments on a wide range of downstream recognition tasks show that our SPT is complementary to the existing PEFT methods.
  arXiv  Detail & Related papers  (2023-03-15T12:34:24Z)
• On the Effectiveness of Parameter-Efficient Fine-Tuning [79.6302606855302]
  Currently, many research works propose to only fine-tune a small portion of the parameters while keeping most of the parameters shared across different tasks. We show that all of the methods are actually sparse fine-tuned models and conduct a novel theoretical analysis of them. Despite the effectiveness of sparsity grounded by our theory, it still remains an open problem of how to choose the tunable parameters.
  arXiv  Detail & Related papers  (2022-11-28T17:41:48Z)
• Know Where You're Going: Meta-Learning for Parameter-Efficient Fine-tuning [34.66092282348687]
  We show that taking the ultimate choice of fine-tuning method into consideration boosts the performance of parameter-efficient fine-tuning. We prime the pretrained model specifically for parameter-efficient fine-tuning, resulting in gains of up to 1.7 points on cross-lingual NER fine-tuning.
  arXiv  Detail & Related papers  (2022-05-25T02:51:57Z)
• No Parameters Left Behind: Sensitivity Guided Adaptive Learning Rate for Training Large Transformer Models [132.90062129639705]
  We propose a novel training strategy that encourages all parameters to be trained sufficiently. A parameter with low sensitivity is redundant, and we improve its fitting by increasing its learning rate. In contrast, a parameter with high sensitivity is well-trained and we regularize it by decreasing its learning rate to prevent further overfitting.
  arXiv  Detail & Related papers  (2022-02-06T00:22:28Z)
• Hyperparameter Selection for Subsampling Bootstraps [0.0]
  A subsampling method like BLB serves as a powerful tool for assessing the quality of estimators for massive data. The performance of the subsampling methods are highly influenced by the selection of tuning parameters. We develop a hyperparameter selection methodology, which can be used to select tuning parameters for subsampling methods. Both simulation studies and real data analysis demonstrate the superior advantage of our method.
  arXiv  Detail & Related papers  (2020-06-02T17:10:45Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.