Initialization of Large Language Models via Reparameterization to Mitigate Loss Spikes

• URL: http://arxiv.org/abs/2410.05052v1
• Date: Mon, 7 Oct 2024 14:09:58 GMT
• Title: Initialization of Large Language Models via Reparameterization to Mitigate Loss Spikes
• Abstract summary: Non-uniformity of the norm of the parameters is one of the causes of loss spikes. In training of neural networks, the scale of the gradients is required to be kept constant throughout the layers.
• Score: 18.12700358872208
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Loss spikes, a phenomenon in which the loss value diverges suddenly, is a fundamental issue in the pre-training of large language models. This paper supposes that the non-uniformity of the norm of the parameters is one of the causes of loss spikes. Here, in training of neural networks, the scale of the gradients is required to be kept constant throughout the layers to avoid the vanishing and exploding gradients problem. However, to meet these requirements in the Transformer model, the norm of the model parameters must be non-uniform, and thus, parameters whose norm is smaller are more sensitive to the parameter update. To address this issue, we propose a novel technique, weight scaling as reparameterization (WeSaR). WeSaR introduces a gate parameter per parameter matrix and adjusts it to the value satisfying the requirements. Because of the gate parameter, WeSaR sets the norm of the original parameters uniformly, which results in stable training. Experimental results with the Transformer decoders consisting of 130 million, 1.3 billion, and 13 billion parameters showed that WeSaR stabilizes and accelerates training and that it outperformed compared methods including popular initialization methods.

Related papers

• Parameterized Representations via Implicit Stochastic Modulation for High-Dimensional and High-Order Neural PDE Solvers [0.24554686192257416]
  Implicit Modulation (PRISM) is a plug-and-play framework for parameterized high-dimensional and high-order neural PDE solvers.<n>We show that it entangles physical parameters with the high-order automatic differentiation, causing extra memory growth and parameter variance amplification.<n>Experiments with PRISM-STDE and PRISM-SDGD on nonlinear parameterized PDEs show stable zero-shot generalization, reduced memory usage, and scalability up to 100,000 dimensions on a single GPU.
  arXiv  Detail & Related papers  (2026-06-20T17:02:30Z)
• FineGates: LLMs Finetuning with Compression using Stochastic Gates [7.093692674858257]
  Large Language Models (LLMs) present significant challenges for full finetuning due to the high computational demands. Lightweight finetuning techniques have been proposed, like learning low-rank adapter layers. We propose an adaptor model based on gates that simultaneously sparsify the frozen base model with task-specific adaptation.
  arXiv  Detail & Related papers  (2024-12-17T14:33:05Z)
• LoRTA: Low Rank Tensor Adaptation of Large Language Models [70.32218116940393]
  Low Rank Adaptation (LoRA) is a popular Efficient Fine Tuning (PEFT) method that effectively adapts large pre-trained models for downstream tasks. We propose a novel approach that employs a low rank tensor parametrization for model updates. Our method is both efficient and effective for fine-tuning large language models, achieving a substantial reduction in the number of parameters while maintaining comparable performance.
  arXiv  Detail & Related papers  (2024-10-05T06:59:50Z)
• Propulsion: Steering LLM with Tiny Fine-Tuning [0.0]
  We propose Propulsion, a novel parameter efficient fine-tuning (PEFT) method to optimize task-specific performance. Inspired by the concept of controlled adjustments in physical motion, Propulsion selectively re-scales specific dimensions of a pre-trained model. Our theoretical analysis, supported by Neural Tangent Kernel (NTK) theory, shows that Propulsion approximates the performance of full fine-tuning with far fewer trainable parameters.
  arXiv  Detail & Related papers  (2024-09-17T06:51:59Z)
• SaRA: High-Efficient Diffusion Model Fine-tuning with Progressive Sparse Low-Rank Adaptation [52.6922833948127]
  In this work, we investigate the importance of parameters in pre-trained diffusion models. We propose a novel model fine-tuning method to make full use of these ineffective parameters. Our method enhances the generative capabilities of pre-trained models in downstream applications.
  arXiv  Detail & Related papers  (2024-09-10T16:44:47Z)
• SVFit: Parameter-Efficient Fine-Tuning of Large Pre-Trained Models Using Singular Values [12.137869917556415]
  Large pre-trained models (LPMs) have demonstrated exceptional performance in diverse natural language processing and computer vision tasks. fully fine-tuning these models poses substantial memory challenges, particularly in resource-constrained environments. We propose SVFit, a novel PEFT approach that leverages singular value decomposition (SVD) to low-rank matrices using critical singular values as trainable parameters.
  arXiv  Detail & Related papers  (2024-09-09T08:44:53Z)
• SMILE: Zero-Shot Sparse Mixture of Low-Rank Experts Construction From Pre-Trained Foundation Models [85.67096251281191]
  We present an innovative approach to model fusion called zero-shot Sparse MIxture of Low-rank Experts (SMILE) construction. SMILE allows for the upscaling of source models into an MoE model without extra data or further training. We conduct extensive experiments across diverse scenarios, such as image classification and text generation tasks, using full fine-tuning and LoRA fine-tuning.
  arXiv  Detail & Related papers  (2024-08-19T17:32:15Z)
• Scaling Exponents Across Parameterizations and Optimizers [94.54718325264218]
  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.
  arXiv  Detail & Related papers  (2024-07-08T12:32:51Z)
• Identifying overparameterization in Quantum Circuit Born Machines [1.7259898169307613]
  We study the onset of over parameterization transitions for quantum circuit Born machines, generative models that are trained using non-adversarial gradient methods. Our results indicate that fully understanding the trainability of these models remains an open question.
  arXiv  Detail & Related papers  (2023-07-06T21:05:22Z)
• 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)
• Regularized Nonlinear Regression for Simultaneously Selecting and Estimating Key Model Parameters [1.6122433144430324]
  In system identification, estimating parameters of a model using limited observations results in poor identifiability. We propose a new method to simultaneously select and estimate sensitive parameters as key model parameters and fix the remaining parameters to a set of typical values.
  arXiv  Detail & Related papers  (2021-04-23T06:17:57Z)
• On the infinite width limit of neural networks with a standard parameterization [52.07828272324366]
  We propose an improved extrapolation of the standard parameterization that preserves all of these properties as width is taken to infinity. We show experimentally that the resulting kernels typically achieve similar accuracy to those resulting from an NTK parameterization.
  arXiv  Detail & Related papers  (2020-01-21T01:02:21Z)

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.