Hoshinori Kanazawa

Authored papers (1)

• Active Inference with a Self-Prior in the Mirror-Mark Task2026

  Spontaneous mark-directed behavior in the mirror-mark task emerges from a single internal mechanism—the self-prior—combined with expected free energy minimization, without any external reward signal. A simulated infant model built on the EMFANT platform (MuJoCo-based, 12 DoF) achieves sticker removal in approximately 70% of evaluation episodes across 80 runs, using only 64×64×3 RGB vision and 12-dimensional proprioception, with tactile input deliberately withheld. The self-prior, implemented as a GPT-like Transformer that autoregressively models joint distributions over 32-class discrete latent variables, is trained exclusively on sticker-free experiences; when a sticker is present, the latent state falls outside the learned high-density region, and the policy—built on STORM's DreamerV3-like world model extended with active inference—selects actions that drive the state back toward familiar self-experience. Across all 80 evaluations, expected free energy after sticker removal (67.33 ± 8.94) was significantly lower than before removal (79.33 ± 4.34; Wilcoxon signed-rank p = 6.33 × 10⁻⁹), and cross-modal sampling confirmed that visual self-appearance can be recovered from proprioception alone, establishing the self-prior as a probabilistic body schema. The paper argues this constitutes a computational implementation of Mitchell's inductive theory of mirror self-recognition and that the free energy principle can serve as a unifying, parsimonious hypothesis for investigating the developmental origins of self-awareness.

More papers — OpenAlex / S2

Recent mentions (1)

• papers-typed

  kim-2026-active-inference.md