Eureka-Moments in Transformers: Multi-Step Tasks Reveal Softmax Induced Optimization Problems

• URL: http://arxiv.org/abs/2310.12956v2
• Date: Thu, 6 Jun 2024 17:58:45 GMT
• Title: Eureka-Moments in Transformers: Multi-Step Tasks Reveal Softmax Induced Optimization Problems
• Authors: David T. Hoffmann, Simon Schrodi, Jelena Bratulić, Nadine Behrmann, Volker Fischer, Thomas Brox,
• Abstract summary: We study rapid improvements of the training loss in transformers when being confronted with multi-step decision tasks. We use synthetic tasks to study the problem in detail, but the leaps in performance can be observed also for language modeling and in-context learning. We find connections and show that ways to improve on the synthetic multi-step tasks can be used to improve the training of language modeling and ICL.
• Score: 27.681141346132286
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: In this work, we study rapid improvements of the training loss in transformers when being confronted with multi-step decision tasks. We found that transformers struggle to learn the intermediate task and both training and validation loss saturate for hundreds of epochs. When transformers finally learn the intermediate task, they do this rapidly and unexpectedly. We call these abrupt improvements Eureka-moments, since the transformer appears to suddenly learn a previously incomprehensible concept. We designed synthetic tasks to study the problem in detail, but the leaps in performance can be observed also for language modeling and in-context learning (ICL). We suspect that these abrupt transitions are caused by the multi-step nature of these tasks. Indeed, we find connections and show that ways to improve on the synthetic multi-step tasks can be used to improve the training of language modeling and ICL. Using the synthetic data we trace the problem back to the Softmax function in the self-attention block of transformers and show ways to alleviate the problem. These fixes reduce the required number of training steps, lead to higher likelihood to learn the intermediate task, to higher final accuracy and training becomes more robust to hyper-parameters.

Related papers

• One-Layer Transformer Provably Learns One-Nearest Neighbor In Context [48.4979348643494]
  We study the capability of one-layer transformers learning the one-nearest neighbor rule. A single softmax attention layer can successfully learn to behave like a one-nearest neighbor.
  arXiv  Detail & Related papers  (2024-11-16T16:12:42Z)
• Dissecting Multiplication in Transformers: Insights into LLMs [23.109124772063574]
  We focus on a typical arithmetic task, integer multiplication, to explore and explain the imperfection of transformers in this domain. We provide comprehensive analysis of a vanilla transformer trained to perform n-digit integer multiplication. We propose improvements to enhance transformers performance on multiplication tasks.
  arXiv  Detail & Related papers  (2024-07-22T04:07:26Z)
• Faith and Fate: Limits of Transformers on Compositionality [109.79516190693415]
  We investigate the limits of transformer large language models across three representative compositional tasks. These tasks require breaking problems down into sub-steps and synthesizing these steps into a precise answer. Our empirical findings suggest that transformer LLMs solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching.
  arXiv  Detail & Related papers  (2023-05-29T23:24:14Z)
• Transformers learn in-context by gradient descent [58.24152335931036]
  Training Transformers on auto-regressive objectives is closely related to gradient-based meta-learning formulations. We show how trained Transformers become mesa-optimizers i.e. learn models by gradient descent in their forward pass.
  arXiv  Detail & Related papers  (2022-12-15T09:21:21Z)
• Don't Sweep your Learning Rate under the Rug: A Closer Look at Cross-modal Transfer of Pretrained Transformers [1.9662978733004601]
  Self-supervised pre-training of large-scale transformer models on text corpora followed by finetuning has achieved state-of-the-art on a number of natural language processing tasks. In our work, we find that this result is, in fact, an artifact of not tuning the learning rates.
  arXiv  Detail & Related papers  (2021-07-26T20:20:48Z)
• A Practical Survey on Faster and Lighter Transformers [0.9176056742068811]
  The Transformer is a model solely based on the attention mechanism that is able to relate any two positions of the input sequence. It has improved the state-of-the-art across numerous sequence modelling tasks. However, its effectiveness comes at the expense of a quadratic computational and memory complexity with respect to the sequence length.
  arXiv  Detail & Related papers  (2021-03-26T17:54:47Z)
• Finetuning Pretrained Transformers into RNNs [81.72974646901136]
  Transformers have outperformed recurrent neural networks (RNNs) in natural language generation. A linear-complexity recurrent variant has proven well suited for autoregressive generation. This work aims to convert a pretrained transformer into its efficient recurrent counterpart.
  arXiv  Detail & Related papers  (2021-03-24T10:50:43Z)
• Continuous Transition: Improving Sample Efficiency for Continuous Control Problems via MixUp [119.69304125647785]
  This paper introduces a concise yet powerful method to construct Continuous Transition. Specifically, we propose to synthesize new transitions for training by linearly interpolating the consecutive transitions. To keep the constructed transitions authentic, we also develop a discriminator to guide the construction process automatically.
  arXiv  Detail & Related papers  (2020-11-30T01:20:23Z)
• Understanding the Difficulty of Training Transformers [120.99980924577787]
  We show that unbalanced gradients are not the root cause of the instability of training. We propose Admin to stabilize the early stage's training and unleash its full potential in the late stage.
  arXiv  Detail & Related papers  (2020-04-17T13:59:07Z)

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.