Towards Empirical Interpretation of Internal Circuits and Properties in Grokked Transformers on Modular Polynomials

• URL: http://arxiv.org/abs/2402.16726v3
• Date: Mon, 18 Nov 2024 02:56:27 GMT
• Title: Towards Empirical Interpretation of Internal Circuits and Properties in Grokked Transformers on Modular Polynomials
• Authors: Hiroki Furuta, Gouki Minegishi, Yusuke Iwasawa, Yutaka Matsuo,
• Abstract summary: Grokking on modular addition has been known to implement Fourier representation and its calculation circuits with trigonometric identities in Transformers. We show that the transferability among the models grokked with each operation can be only limited to specific combinations. Some multi-task mixtures may lead to co-grokking, where grokking simultaneously happens for all the tasks.
• Score: 29.09237503747052
• License:
• Abstract: Grokking has been actively explored to reveal the mystery of delayed generalization and identifying interpretable representations and algorithms inside the grokked models is a suggestive hint to understanding its mechanism. Grokking on modular addition has been known to implement Fourier representation and its calculation circuits with trigonometric identities in Transformers. Considering the periodicity in modular arithmetic, the natural question is to what extent these explanations and interpretations hold for the grokking on other modular operations beyond addition. For a closer look, we first hypothesize that any modular operations can be characterized with distinctive Fourier representation or internal circuits, grokked models obtain common features transferable among similar operations, and mixing datasets with similar operations promotes grokking. Then, we extensively examine them by learning Transformers on complex modular arithmetic tasks, including polynomials. Our Fourier analysis and novel progress measure for modular arithmetic, Fourier Frequency Density and Fourier Coefficient Ratio, characterize distinctive internal representations of grokked models per modular operation; for instance, polynomials often result in the superposition of the Fourier components seen in elementary arithmetic, but clear patterns do not emerge in challenging non-factorizable polynomials. In contrast, our ablation study on the pre-grokked models reveals that the transferability among the models grokked with each operation can be only limited to specific combinations, such as from elementary arithmetic to linear expressions. Moreover, some multi-task mixtures may lead to co-grokking -- where grokking simultaneously happens for all the tasks -- and accelerate generalization, while others may not find optimal solutions. We provide empirical steps towards the interpretability of internal circuits.

Related papers

• Learning Linear Attention in Polynomial Time [115.68795790532289]
  We provide the first results on learnability of single-layer Transformers with linear attention. We show that linear attention may be viewed as a linear predictor in a suitably defined RKHS. We show how to efficiently identify training datasets for which every empirical riskr is equivalent to the linear Transformer.
  arXiv  Detail & Related papers  (2024-10-14T02:41:01Z)
• Generalization of Modular Spread Complexity for Non-Hermitian Density Matrices [0.0]
  In this work we generalize the concept of modular spread complexity to the cases where the reduced density matrix is non-Hermitian. We define the quantity pseudo-capacity which generalizes capacity of entanglement, and corresponds to the early modular-time measure of pseudo-modular complexity. We show some analytical calculations for 2-level systems and 4-qubit models and then do numerical investigations on the quantum phase transition of transverse field Ising model.
  arXiv  Detail & Related papers  (2024-10-07T17:59:16Z)
• In-Context Learning with Representations: Contextual Generalization of Trained Transformers [66.78052387054593]
  In-context learning (ICL) refers to a capability of pretrained large language models, which can learn a new task given a few examples during inference. This paper investigates the training dynamics of transformers by gradient descent through the lens of non-linear regression tasks.
  arXiv  Detail & Related papers  (2024-08-19T16:47:46Z)
• Shape Arithmetic Expressions: Advancing Scientific Discovery Beyond Closed-Form Equations [56.78271181959529]
  Generalized Additive Models (GAMs) can capture non-linear relationships between variables and targets, but they cannot capture intricate feature interactions. We propose Shape Expressions Arithmetic ( SHAREs) that fuses GAM's flexible shape functions with the complex feature interactions found in mathematical expressions. We also design a set of rules for constructing SHAREs that guarantee transparency of the found expressions beyond the standard constraints.
  arXiv  Detail & Related papers  (2024-04-15T13:44:01Z)
• Discovering modular solutions that generalize compositionally [55.46688816816882]
  We show that identification up to linear transformation purely from demonstrations is possible without having to learn an exponential number of module combinations. We further demonstrate empirically that meta-learning from finite data can discover modular policies that generalize compositionally in a number of complex environments.
  arXiv  Detail & Related papers  (2023-12-22T16:33:50Z)
• DIFFormer: Scalable (Graph) Transformers Induced by Energy Constrained Diffusion [66.21290235237808]
  We introduce an energy constrained diffusion model which encodes a batch of instances from a dataset into evolutionary states. We provide rigorous theory that implies closed-form optimal estimates for the pairwise diffusion strength among arbitrary instance pairs. Experiments highlight the wide applicability of our model as a general-purpose encoder backbone with superior performance in various tasks.
  arXiv  Detail & Related papers  (2023-01-23T15:18:54Z)
• High-Dimensional Undirected Graphical Models for Arbitrary Mixed Data [2.2871867623460207]
  In many applications data span variables of different types, whose principled joint analysis is nontrivial. Recent advances have shown how the binary-continuous case can be tackled, but the general mixed variable type regime remains challenging. We propose flexible and scalable methodology for data with variables of entirely general mixed type.
  arXiv  Detail & Related papers  (2022-11-21T18:21:31Z)
• Inductive Biases and Variable Creation in Self-Attention Mechanisms [25.79946667926312]
  This work provides a theoretical analysis of the inductive biases of self-attention modules. Our focus is to rigorously establish which functions and long-range dependencies self-attention blocks prefer to represent. Our main result shows that bounded-norm Transformer layers create sparse variables.
  arXiv  Detail & Related papers  (2021-10-19T16:36:19Z)
• A Compositional Atlas of Tractable Circuit Operations: From Simple Transformations to Complex Information-Theoretic Queries [44.36335714431731]
  We show how complex inference scenarios for machine learning can be represented in terms of tractable modular operations over circuits. We derive a unified framework for reasoning about tractable models that generalizes several results in the literature and opens up novel tractable inference scenarios.
  arXiv  Detail & Related papers  (2021-02-11T17:26:32Z)
• A Deep Joint Sparse Non-negative Matrix Factorization Framework for Identifying the Common and Subject-specific Functional Units of Tongue Motion During Speech [7.870139900799612]
  We develop a new deep learning framework to identify common and subject-specific functional units of tongue motion during speech. We transform NMF with sparse and graph regularizations into modular architectures akin to deep neural networks.
  arXiv  Detail & Related papers  (2020-07-09T15:05:44Z)
• From Sets to Multisets: Provable Variational Inference for Probabilistic Integer Submodular Models [82.95892656532696]
  Submodular functions have been studied extensively in machine learning and data mining. In this work, we propose a continuous DR-submodular extension for integer submodular functions. We formulate a new probabilistic model which is defined through integer submodular functions.
  arXiv  Detail & Related papers  (2020-06-01T22:20:45Z)

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.