[1Yin-A] Poster 1

[1Yin-A-56] Interpretable Neuro-Symbolic Inference for Analog Clock Time Recognition with DeepProbLog

〇Ruka Maruyama¹, Takashi Kaburagi¹ (1. International Christian University)

Keywords:

AI,Neuro-Symbolic,Inference

This paper proposes a neuro-symbolic approach to analog clock time recognition using DeepProbLog with latent variable learning. Our method decomposes the clock-reading task into three latent components, namely dial rotation, hour-hand angle, and minute-hand angle, which are learned solely from time labels without geometric annotations. A shared EfficientNet-B0 backbone feeds three neural predicate heads whose outputs are constrained by a ProbLog logic program encoding rotation correction, hour-minute coupling, and tolerance-based matching. On a Kaggle clock dataset with 1,440 test images, our full model achieves 74.58% exact time accuracy, while a CNN-only baseline attains higher raw accuracy (82.78%), it provides no decomposition into interpretable components. We show that our model learns meaningful latent representations, achieving 52.08% rotation, 39.03% hour-hand, and 58.06% minute-hand accuracy against geometric ground truth, which demonstrates that logical constraints effectively guide latent variable learning toward physically grounded representations.