Explainable Artificial Intelligence

AIME²

Toward a unified algebraic theory of explainability via approximate inverse operators

AIME² formulates explainability through an approximate inverse-operator perspective and introduces an Explainability Operator for studying global, output-wise, class-wise, and local explanations within a common algebraic framework.

About

Takafumi Nakanishi

Professor at the School of Computer Science, Tokyo University of Technology. My research focuses on explainable artificial intelligence, inverse-operator approaches to explanation, data mining, and mathematical models for media and intelligent systems.

I proposed Approximate Inverse Model Explanations (AIME) and continue to develop its theoretical foundations, robust and uncertainty-aware extensions, software, and applications.

AIME

Approximate Inverse Model Explanations

AIME is a model-agnostic XAI framework that learns an approximate inverse operator from observed model behavior and uses it to construct input-side explanatory structures.

The original AIME paper introduced the method and demonstrated local and global insights across multiple data modalities. AIME² develops the broader algebraic foundation of the Explainability Operator.

AIME²

From attribution outputs to an Explainability Operator

AIME² studies explanation as an operator connecting model-output representations to explanatory structures in the input space. It does not claim that all XAI methods are identical; it provides an algebraic framework for analyzing a family of inverse-operator explanations.

Forward prediction Inputs are mapped to model-output representations.
Explainability Operator A weighted and regularized approximate inverse connects output-side information to input-side explanatory structures.
Explanation queries Related explanatory objects are evaluated within one framework.
GlobalOutput-wiseClass-wiseLocal

Conceptual overview of the AIME² inverse-operator perspective.

Common algebraic basis

Multiple explanation queries can be studied as related evaluations of an operator-based framework.

Weighted formulation

Weighting changes the analytical scope and supports globally or locally emphasized explanations.

Regularized estimation

Regularization provides a principled response to ill-conditioning and instability in approximate inversion.

Algebraic analysis

The framework analyzes reconstruction, minimum-norm solutions, stability, and coordinate equivariance.

Publications

When should you cite AIME²?

Please consider citing AIME² when your research addresses one or more of the following:

  • explainability formulated through an inverse or reconstruction problem;
  • explanations represented as operators rather than isolated attribution vectors;
  • a common mathematical basis for global, output-wise, class-wise, and local explanations;
  • generalized inverses, minimum-norm solutions, weighting, or regularization in XAI;
  • reconstruction, stability, or coordinate equivariance of explanations;
  • the relationship between forward prediction and inverse explanation;
  • model auditing or repeated explanation queries based on a reusable explanation operator; or
  • theoretical comparisons between attribution vectors and operator-valued explanations.

In one sentence:
Cite AIME² when you need an algebraic foundation for studying explainability through a weighted and regularized approximate inverse operator.

Article

T. Nakanishi, “AIME2: toward a unified algebraic theory of explainability via approximate inverse operators,” Communications AI & Computing, vol. 1, no. 1, Art. no. 3, 2026.

DOI: 10.1038/s44488-026-00004-0

BibTeX

View the full publication list →

Software

Articles, code, and implementation resources

Contact

Research collaboration and discussion

For academic discussion, collaboration, or questions about AIME and AIME², please contact me through my official website or GitHub profile.