Talking Heads: Understanding Inter-layer Communication in Transformer Language Models

• URL: http://arxiv.org/abs/2406.09519v2
• Date: Sun, 03 Nov 2024 17:48:48 GMT
• Title: Talking Heads: Understanding Inter-layer Communication in Transformer Language Models
• Authors: Jack Merullo, Carsten Eickhoff, Ellie Pavlick,
• Abstract summary: We analyze a mechanism used in two LMs to selectively inhibit items in a context in one task. We find that models write into low-rank subspaces of the residual stream to represent features which are then read out by later layers.
• Score: 32.2976613483151
• License:
• Abstract: Although it is known that transformer language models (LMs) pass features from early layers to later layers, it is not well understood how this information is represented and routed by the model. We analyze a mechanism used in two LMs to selectively inhibit items in a context in one task, and find that it underlies a commonly used abstraction across many context-retrieval behaviors. Specifically, we find that models write into low-rank subspaces of the residual stream to represent features which are then read out by later layers, forming low-rank communication channels (Elhage et al., 2021) between layers. A particular 3D subspace in model activations in GPT-2 can be traversed to positionally index items in lists, and we show that this mechanism can explain an otherwise arbitrary-seeming sensitivity of the model to the order of items in the prompt. That is, the model has trouble copying the correct information from context when many items ``crowd" this limited space. By decomposing attention heads with the Singular Value Decomposition (SVD), we find that previously described interactions between heads separated by one or more layers can be predicted via analysis of their weight matrices alone. We show that it is possible to manipulate the internal model representations as well as edit model weights based on the mechanism we discover in order to significantly improve performance on our synthetic Laundry List task, which requires recall from a list, often improving task accuracy by over 20%. Our analysis reveals a surprisingly intricate interpretable structure learned from language model pretraining, and helps us understand why sophisticated LMs sometimes fail in simple domains, facilitating future analysis of more complex behaviors.

Related papers

• Two are better than one: Context window extension with multi-grained self-injection [111.1376461868317]
  SharedLLM is a novel approach grounded in the design philosophy of multi-grained context compression and query-aware information retrieval. We introduce a specialized tree-style data structure to efficiently encode, store and retrieve multi-grained contextual information for text chunks.
  arXiv  Detail & Related papers  (2024-10-25T06:08:59Z)
• The Mysterious Case of Neuron 1512: Injectable Realignment Architectures Reveal Internal Characteristics of Meta's Llama 2 Model [3.838217057990932]
  Injectable Realignment Model (IRM) is a novel approach to language model interpretability and explainability. Inspired by earlier work on Neural Programming Interfaces, we construct and train a small network -- the IRM -- to induce emotion-based alignments. Analysis of the trained IRM's outputs reveals a curious pattern.
  arXiv  Detail & Related papers  (2024-07-04T04:05:19Z)
• Explaining Text Similarity in Transformer Models [52.571158418102584]
  Recent advances in explainable AI have made it possible to mitigate limitations by leveraging improved explanations for Transformers. We use BiLRP, an extension developed for computing second-order explanations in bilinear similarity models, to investigate which feature interactions drive similarity in NLP models. Our findings contribute to a deeper understanding of different semantic similarity tasks and models, highlighting how novel explainable AI methods enable in-depth analyses and corpus-level insights.
  arXiv  Detail & Related papers  (2024-05-10T17:11:31Z)
• In-Context Language Learning: Architectures and Algorithms [73.93205821154605]
  We study ICL through the lens of a new family of model problems we term in context language learning (ICLL) We evaluate a diverse set of neural sequence models on regular ICLL tasks.
  arXiv  Detail & Related papers  (2024-01-23T18:59:21Z)
• Language Models Implement Simple Word2Vec-style Vector Arithmetic [32.2976613483151]
  A primary criticism towards language models (LMs) is their inscrutability. This paper presents evidence that, despite their size and complexity, LMs sometimes exploit a simple vector arithmetic style mechanism to solve some relational tasks.
  arXiv  Detail & Related papers  (2023-05-25T15:04:01Z)
• Physics of Language Models: Part 1, Learning Hierarchical Language Structures [51.68385617116854]
  Transformer-based language models are effective but complex, and understanding their inner workings is a significant challenge. We introduce a family of synthetic CFGs that produce hierarchical rules, capable of generating lengthy sentences. We demonstrate that generative models like GPT can accurately learn this CFG language and generate sentences based on it.
  arXiv  Detail & Related papers  (2023-05-23T04:28:16Z)
• VISIT: Visualizing and Interpreting the Semantic Information Flow of Transformers [45.42482446288144]
  Recent advances in interpretability suggest we can project weights and hidden states of transformer-based language models to their vocabulary. We investigate LM attention heads and memory values, the vectors the models dynamically create and recall while processing a given input. We create a tool to visualize a forward pass of Generative Pre-trained Transformers (GPTs) as an interactive flow graph.
  arXiv  Detail & Related papers  (2023-05-22T19:04:56Z)
• Interpretability at Scale: Identifying Causal Mechanisms in Alpaca [62.65877150123775]
  We use Boundless DAS to efficiently search for interpretable causal structure in large language models while they follow instructions. Our findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models.
  arXiv  Detail & Related papers  (2023-05-15T17:15:40Z)
• ELUDE: Generating interpretable explanations via a decomposition into labelled and unlabelled features [23.384134043048807]
  We develop an explanation framework that decomposes a model's prediction into two parts. By identifying the latter, we are able to analyze the "unexplained" portion of the model. We show that the set of unlabelled features can generalize to multiple models trained with the same feature space.
  arXiv  Detail & Related papers  (2022-06-15T17:36:55Z)
• Learning Contextual Representations for Semantic Parsing with Generation-Augmented Pre-Training [86.91380874390778]
  We present Generation-Augmented Pre-training (GAP), that jointly learns representations of natural language utterances and table schemas by leveraging generation models to generate pre-train data. Based on experimental results, neural semantics that leverage GAP MODEL obtain new state-of-the-art results on both SPIDER and CRITERIA-TO-generative benchmarks.
  arXiv  Detail & Related papers  (2020-12-18T15:53:50Z)

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.