Transcoders Find Interpretable LLM Feature Circuits

• URL: http://arxiv.org/abs/2406.11944v2
• Date: Wed, 06 Nov 2024 22:37:30 GMT
• Title: Transcoders Find Interpretable LLM Feature Circuits
• Authors: Jacob Dunefsky, Philippe Chlenski, Neel Nanda,
• Abstract summary: We introduce a novel method for using transcoders to perform circuit analysis through sublayers. We train transcoders on language models with 120M, 410M, and 1.4B parameters, and find them to perform at least on par with SAEs in terms of sparsity, faithfulness, and human-interpretability.
• Score: 1.4254279830438588
• License:
• Abstract: A key goal in mechanistic interpretability is circuit analysis: finding sparse subgraphs of models corresponding to specific behaviors or capabilities. However, MLP sublayers make fine-grained circuit analysis on transformer-based language models difficult. In particular, interpretable features -- such as those found by sparse autoencoders (SAEs) -- are typically linear combinations of extremely many neurons, each with its own nonlinearity to account for. Circuit analysis in this setting thus either yields intractably large circuits or fails to disentangle local and global behavior. To address this we explore transcoders, which seek to faithfully approximate a densely activating MLP layer with a wider, sparsely-activating MLP layer. We introduce a novel method for using transcoders to perform weights-based circuit analysis through MLP sublayers. The resulting circuits neatly factorize into input-dependent and input-invariant terms. We then successfully train transcoders on language models with 120M, 410M, and 1.4B parameters, and find them to perform at least on par with SAEs in terms of sparsity, faithfulness, and human-interpretability. Finally, we apply transcoders to reverse-engineer unknown circuits in the model, and we obtain novel insights regarding the "greater-than circuit" in GPT2-small. Our results suggest that transcoders can prove effective in decomposing model computations involving MLPs into interpretable circuits. Code is available at https://github.com/jacobdunefsky/transcoder_circuits/.

Related papers

• LaMAGIC: Language-Model-based Topology Generation for Analog Integrated Circuits [17.002169206594793]
  We introduce LaMAGIC, a pioneering language model-based topology generation model. LaMAGIC can efficiently generate an optimized circuit design from the custom specification in a single pass. LaMAGIC achieves a success rate of up to 96% under a strict tolerance of 0.01.
  arXiv  Detail & Related papers  (2024-07-19T22:51:41Z)
• Interpreting Attention Layer Outputs with Sparse Autoencoders [3.201633659481912]
  Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work.
  arXiv  Detail & Related papers  (2024-06-25T17:43:13Z)
• Finding Transformer Circuits with Edge Pruning [71.12127707678961]
  We propose Edge Pruning as an effective and scalable solution to automated circuit discovery. Our method finds circuits in GPT-2 that use less than half the number of edges compared to circuits found by previous methods. Thanks to its efficiency, we scale Edge Pruning to CodeLlama-13B, a model over 100x the scale that prior methods operate on.
  arXiv  Detail & Related papers  (2024-06-24T16:40:54Z)
• Transformers need glasses! Information over-squashing in language tasks [18.81066657470662]
  We study how information propagates in decoder-only Transformers. We show that certain sequences of inputs to the Transformer can yield arbitrarily close representations in the final token. We also show that decoder-only Transformer language models can lose sensitivity to specific tokens in the input.
  arXiv  Detail & Related papers  (2024-06-06T17:14:44Z)
• Automatically Identifying Local and Global Circuits with Linear Computation Graphs [45.760716193942685]
  We introduce our circuit discovery pipeline with Sparse Autoencoders (SAEs) and a variant called Transcoders. Our methods do not require linear approximation to compute the causal effect of each node. We analyze three kinds of circuits in GPT-2 Small: bracket, induction, and Indirect Object Identification circuits.
  arXiv  Detail & Related papers  (2024-05-22T17:50:04Z)
• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models [55.19497659895122]
  We introduce methods for discovering and applying sparse feature circuits. These are causally implicatedworks of human-interpretable features for explaining language model behaviors.
  arXiv  Detail & Related papers  (2024-03-28T17:56:07Z)
• Circuit Transformer: End-to-end Circuit Design by Predicting the Next Gate [20.8279111910994]
  Language, a prominent human ability to express through sequential symbols, has been computationally mastered by recent advances of large language models (LLMs) LLMs have shown unprecedented capabilities in understanding and reasoning. Can circuits also be mastered by a a sufficiently large "circuit model", which can conquer electronic design tasks by simply predicting the next logic gate?
  arXiv  Detail & Related papers  (2024-03-14T03:24:14Z)
• Machine Learning-Aided Efficient Decoding of Reed-Muller Subcodes [59.55193427277134]
  Reed-Muller (RM) codes achieve the capacity of general binary-input memoryless symmetric channels. RM codes only admit limited sets of rates. Efficient decoders are available for RM codes at finite lengths.
  arXiv  Detail & Related papers  (2023-01-16T04:11:14Z)
• Transformers Learn Shortcuts to Automata [52.015990420075944]
  We find that a low-depth Transformer can represent the computations of any finite-state automaton. We show that a Transformer with $O(log T)$ layers can exactly replicate the computation of an automaton on an input sequence of length $T$. We further investigate the brittleness of these solutions and propose potential mitigations.
  arXiv  Detail & Related papers  (2022-10-19T17:45:48Z)
• Learning to Encode Position for Transformer with Continuous Dynamical Model [88.69870971415591]
  We introduce a new way of learning to encode position information for non-recurrent models, such as Transformer models. We model the evolution of encoded results along position index by such a dynamical system.
  arXiv  Detail & Related papers  (2020-03-13T00:41:41Z)

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.