Initialization is Critical to Whether Transformers Fit Composite Functions by Inference or Memorizing

• URL: http://arxiv.org/abs/2405.05409v3
• Date: Sat, 05 Oct 2024 16:31:26 GMT
• Title: Initialization is Critical to Whether Transformers Fit Composite Functions by Inference or Memorizing
• Authors: Zhongwang Zhang, Pengxiao Lin, Zhiwei Wang, Yaoyu Zhang, Zhi-Qin John Xu,
• Abstract summary: Transformers have shown impressive capabilities across various tasks, but their performance on compositional problems remains a topic of debate. In this work, we investigate the mechanisms of how transformers behave on unseen compositional tasks.
• Score: 10.206921909332006
• License:
• Abstract: Transformers have shown impressive capabilities across various tasks, but their performance on compositional problems remains a topic of debate. In this work, we investigate the mechanisms of how transformers behave on unseen compositional tasks. We discover that the parameter initialization scale plays a critical role in determining whether the model learns inferential solutions, which capture the underlying compositional primitives, or symmetric solutions, which simply memorize mappings without understanding the compositional structure. By analyzing the information flow and vector representations within the model, we reveal the distinct mechanisms underlying these solution types. We further find that inferential solutions exhibit low complexity bias, which we hypothesize is a key factor enabling them to learn individual mappings for single anchors. We validate our conclusions on various real-world datasets. Our findings provide valuable insights into the role of initialization scale in shaping the type of solution learned by transformers and their ability to learn and generalize compositional tasks.

Related papers

• Counting in Small Transformers: The Delicate Interplay between Attention and Feed-Forward Layers [16.26331213222281]
  We investigate how architectural design choices influence the space of solutions that a transformer can implement and learn. We characterize two different counting strategies that small transformers can implement theoretically. Our findings highlight that even in simple settings, slight variations in model design can cause significant changes to the solutions a transformer learns.
  arXiv  Detail & Related papers  (2024-07-16T09:48:10Z)
• Transformers are uninterpretable with myopic methods: a case study with bounded Dyck grammars [36.780346257061495]
  Interpretability methods aim to understand the algorithm implemented by a trained model. We take a critical view of methods that exclusively focus on individual parts of the model.
  arXiv  Detail & Related papers  (2023-12-03T15:34:46Z)
• How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations [98.7450564309923]
  This paper takes initial steps on understanding in-context learning (ICL) in more complex scenarios, by studying learning with representations. We construct synthetic in-context learning problems with a compositional structure, where the label depends on the input through a possibly complex but fixed representation function. We show theoretically the existence of transformers that approximately implement such algorithms with mild depth and size.
  arXiv  Detail & Related papers  (2023-10-16T17:40:49Z)
• From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication [19.336940758147442]
  It has been observed that representations learned by distinct neural networks conceal structural similarities when the models are trained under similar inductive biases. We introduce a versatile method to directly incorporate a set of invariances into the representations, constructing a product space of invariant components on top of the latent representations. We validate our solution on classification and reconstruction tasks, observing consistent latent similarity and downstream performance improvements in a zero-shot stitching setting.
  arXiv  Detail & Related papers  (2023-10-02T13:55:38Z)
• Leveraging sparse and shared feature activations for disentangled representation learning [112.22699167017471]
  We propose to leverage knowledge extracted from a diversified set of supervised tasks to learn a common disentangled representation. We validate our approach on six real world distribution shift benchmarks, and different data modalities.
  arXiv  Detail & Related papers  (2023-04-17T01:33:24Z)
• Unveiling Transformers with LEGO: a synthetic reasoning task [23.535488809197787]
  We study how the transformer architecture learns to follow a chain of reasoning. In some data regime the trained transformer finds "shortcut" solutions to follow the chain of reasoning. We find that one can prevent such shortcut with appropriate architecture modification or careful data preparation.
  arXiv  Detail & Related papers  (2022-06-09T06:30:17Z)
• Transformer for Partial Differential Equations' Operator Learning [0.0]
  We present an attention-based framework for data-driven operator learning, which we term Operator Transformer (OFormer) Our framework is built upon self-attention, cross-attention, and a set of point-wise multilayer perceptrons (MLPs)
  arXiv  Detail & Related papers  (2022-05-26T23:17:53Z)
• Amortized Inference for Causal Structure Learning [72.84105256353801]
  Learning causal structure poses a search problem that typically involves evaluating structures using a score or independence test. We train a variational inference model to predict the causal structure from observational/interventional data. Our models exhibit robust generalization capabilities under substantial distribution shift.
  arXiv  Detail & Related papers  (2022-05-25T17:37:08Z)
• Eigen Analysis of Self-Attention and its Reconstruction from Partial Computation [58.80806716024701]
  We study the global structure of attention scores computed using dot-product based self-attention. We find that most of the variation among attention scores lie in a low-dimensional eigenspace. We propose to compute scores only for a partial subset of token pairs, and use them to estimate scores for the remaining pairs.
  arXiv  Detail & Related papers  (2021-06-16T14:38:42Z)
• Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision [57.14468881854616]
  We propose an auxiliary training objective that improves the generalization capabilities of neural networks. We use pairs of minimally-different examples with different labels, a.k.a counterfactual or contrasting examples, which provide a signal indicative of the underlying causal structure of the task. Models trained with this technique demonstrate improved performance on out-of-distribution test sets.
  arXiv  Detail & Related papers  (2020-04-20T02:47:49Z)
• Invariant Feature Coding using Tensor Product Representation [75.62232699377877]
  We prove that the group-invariant feature vector contains sufficient discriminative information when learning a linear classifier. A novel feature model that explicitly consider group action is proposed for principal component analysis and k-means clustering.
  arXiv  Detail & Related papers  (2019-06-05T07:15:17Z)

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.