Why do small language models underperform? Studying Language Model Saturation via the Softmax Bottleneck

• URL: http://arxiv.org/abs/2404.07647v1
• Date: Thu, 11 Apr 2024 11:10:36 GMT
• Title: Why do small language models underperform? Studying Language Model Saturation via the Softmax Bottleneck
• Authors: Nathan Godey, Éric de la Clergerie, Benoît Sagot,
• Abstract summary: We find that smaller models can suffer from saturation, characterized as a drop in performance at some advanced point in training followed by a plateau. This can be explained by a mismatch between the hidden dimension of smaller models and the high rank of the target contextual probability distribution. We measure the effect of the softmax bottleneck in various settings and find that models based on less than 1000 hidden dimensions tend to adopt degenerate latent representations in late pretraining.
• Score: 11.416426888383873
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Recent advances in language modeling consist in pretraining highly parameterized neural networks on extremely large web-mined text corpora. Training and inference with such models can be costly in practice, which incentivizes the use of smaller counterparts. However, it has been observed that smaller models can suffer from saturation, characterized as a drop in performance at some advanced point in training followed by a plateau. In this paper, we find that such saturation can be explained by a mismatch between the hidden dimension of smaller models and the high rank of the target contextual probability distribution. This mismatch affects the performance of the linear prediction head used in such models through the well-known softmax bottleneck phenomenon. We measure the effect of the softmax bottleneck in various settings and find that models based on less than 1000 hidden dimensions tend to adopt degenerate latent representations in late pretraining, which leads to reduced evaluation performance.

Related papers

• A Hitchhiker's Guide to Scaling Law Estimation [56.06982415792523]
  Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. We estimate more than 1000 scaling laws, then derive a set of best practices for estimating scaling laws in new model families.
  arXiv  Detail & Related papers  (2024-10-15T17:59:10Z)
• Tending Towards Stability: Convergence Challenges in Small Language Models [3.734405405403176]
  Despite their advantages, smaller models frequently underperform compared to their larger counterparts. This is anecdotally attributed to their reduced representational capacity. By linking the convergence of layers' activations to their parameters' effective rank, our analyses can guide future work to address inefficiencies in the learning dynamics of small models.
  arXiv  Detail & Related papers  (2024-10-15T09:57:19Z)
• Strong Model Collapse [16.071600606637908]
  We consider a supervised regression setting and establish the existance of a strong form of the model collapse phenomenon. Our results show that even the smallest fraction of synthetic data can lead to model collapse. We investigate whether increasing model size, an approach aligned with current trends in training large language models, exacerbates or mitigates model collapse.
  arXiv  Detail & Related papers  (2024-10-07T08:54:23Z)
• Causal Estimation of Memorisation Profiles [58.20086589761273]
  Understanding memorisation in language models has practical and societal implications. Memorisation is the causal effect of training with an instance on the model's ability to predict that instance. This paper proposes a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics.
  arXiv  Detail & Related papers  (2024-06-06T17:59:09Z)
• A Dynamical Model of Neural Scaling Laws [79.59705237659547]
  We analyze a random feature model trained with gradient descent as a solvable model of network training and generalization. Our theory shows how the gap between training and test loss can gradually build up over time due to repeated reuse of data.
  arXiv  Detail & Related papers  (2024-02-02T01:41:38Z)
• Compressing Sentence Representation with maximum Coding Rate Reduction [0.0]
  In most natural language inference problems, sentence representation is needed for semantic retrieval tasks. Due to space and time hardware limitations, there is a need to attain comparable results when using the smaller model. We demonstrate that the new language model with reduced complexity and sentence embedding size can achieve comparable results on semantic retrieval benchmarks.
  arXiv  Detail & Related papers  (2023-04-25T09:23:43Z)
• Training Trajectories of Language Models Across Scales [99.38721327771208]
  Scaling up language models has led to unprecedented performance gains. How do language models of different sizes learn during pre-training? Why do larger language models demonstrate more desirable behaviors?
  arXiv  Detail & Related papers  (2022-12-19T19:16:29Z)
• A Multi-dimensional Evaluation of Tokenizer-free Multilingual Pretrained Models [87.7086269902562]
  We show that subword-based models might still be the most practical choice in many settings. We encourage future work in tokenizer-free methods to consider these factors when designing and evaluating new models.
  arXiv  Detail & Related papers  (2022-10-13T15:47:09Z)
• A Bayesian Perspective on Training Speed and Model Selection [51.15664724311443]
  We show that a measure of a model's training speed can be used to estimate its marginal likelihood. We verify our results in model selection tasks for linear models and for the infinite-width limit of deep neural networks. Our results suggest a promising new direction towards explaining why neural networks trained with gradient descent are biased towards functions that generalize well.
  arXiv  Detail & Related papers  (2020-10-27T17:56:14Z)
• On the Benefits of Models with Perceptually-Aligned Gradients [8.427953227125148]
  We show that interpretable and perceptually aligned gradients are present even in models that do not show high robustness to adversarial attacks. We leverage models with interpretable perceptually-aligned features and show that adversarial training with low max-perturbation bound can improve the performance of models for zero-shot and weakly supervised localization tasks.
  arXiv  Detail & Related papers  (2020-05-04T14:05:38Z)

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.