When Rule Violations Are Rare: Chimera Training for Logical Anomaly Detection

Many practical anomalies are not merely rare inputs, but violations of semantic constraints: objects co-occur in structured ways, actions imply preconditions, and events satisfy temporal or relational regularities. We study anomaly detection in this setting, where constraints are given as logical rules over learned visual concepts, but real rule violations are rare or absent during training. We propose a neural rule evaluator that compiles each constraint into a directed acyclic graph and learns feature-aware subtree MLP gates for its internal logical operators. Each gate maps child features and edge-level negations to a parent representation and a rule-satisfaction probability, with intermediate supervision obtained from exact Boolean propagation over ground-truth concept labels. The key difficulty is that same-image training data often provide insufficient coverage of informative truth configurations and also allow shortcut solutions. To address this, we introduce chimera training: an operand-level counterfactual construction at the feature level. Instead of mixing input images, we concatenate subtree features from different samples; each operand keeps the hard truth label of the sample it came from, and the chimera target is obtained by applying the node's logical operator to those inherited labels. This supplies supervised logical counterexamples without requiring real anomalous images. Across CLEVRER, OpenImages, and VidOR, the resulting evaluator improves rule-level anomaly AUROC over independent-events and same-image semantic-training baselines, especially for compositional and relational rules. The method yields both scalar anomaly scores and rule-level attributions.

📋 한 줄 요약

**[Logical Anomaly Detection / Chimera Training]** Chimera training이 operand-level counterfactual feature concatenation으로 rule violation 데이터 없이 logical anomaly detector 학습, CLEVRER·OpenImages·VidOR에서 compositional·relational rule AUROC 개선.

🎯 핵심 기여도

많은 실용 anomaly가 단순 rare input이 아니라 semantic constraint 위반임을 강조 — object는 structured하게 co-occur, action은 precondition을 imply, event는 temporal·relational regularity를 만족.
Constraint가 learned visual concept 위 logical rule로 주어지지만 real rule violation이 학습 중 rare·absent한 환경에서의 anomaly detection 연구.
Neural rule evaluator 제안 — 각 constraint를 directed acyclic graph로 컴파일, internal logical operator마다 feature-aware subtree MLP gate 학습 — gate가 child feature·edge-level negation을 parent representation과 rule-satisfaction probability에 매핑, ground-truth concept label에 대한 exact Boolean propagation으로 intermediate supervision 획득.
핵심 난제 — same-image training data가 종종 informative truth configuration의 coverage 부족·shortcut solution 허용 — Chimera training으로 해결: feature 수준의 operand-level counterfactual 구성으로 입력 이미지 mixing 대신 다른 sample의 subtree feature를 concatenate, 각 operand는 출신 sample의 hard truth label 유지, chimera target은 inherited label에 노드의 logical operator 적용으로 획득 — real anomalous image 없이도 supervised logical counterexample 공급.

💡 핵심 아이디어

Logical anomaly detection의 핵심 난제는 real rule violation 데이터 부재이며, image-level mixing이 아닌 feature-level operand 수준에서 다른 sample의 subtree feature를 concatenate해 hard truth label과 operator로 chimera target을 자동 합성하면, 실제 anomalous image 없이도 supervised logical counterexample 학습이 가능하다.

🔬 기술적 접근법

**방법론**: Neural rule evaluator + Chimera training.
**핵심 기법**: (1) 각 constraint를 directed acyclic graph로 컴파일, (2) Internal logical operator에 feature-aware subtree MLP gate 학습 — child feature·edge negation을 parent representation·rule-satisfaction probability에 매핑, (3) Ground-truth concept label의 exact Boolean propagation으로 intermediate supervision, (4) Chimera training — feature-level operand counterfactual로 다른 sample의 subtree feature concatenate, hard truth label 유지·operator 적용으로 chimera target 합성, (5) Scalar anomaly score와 rule-level attribution 동시 산출.

📊 주요 결과

CLEVRER·OpenImages·VidOR 3 데이터셋에서 rule-level anomaly AUROC가 independent-events·same-image semantic-training baseline 대비 개선.
특히 compositional·relational rule에서 큰 향상.
Scalar anomaly score와 rule-level attribution 양쪽 산출.
Real anomalous image 없이 효과적 학습 가능.

💭 의의 및 한계

**의의**: Logical anomaly detection의 데이터 부족 문제를 feature-level operand chimera로 우회, neurosymbolic 학습의 새 데이터 augmentation 패러다임, 3 다양 도메인 데이터셋에서 일반화 입증, rule-level attribution으로 interpretable. **한계**: Logical rule이 사전 주어져야 함, learned visual concept의 정확도에 종속, chimera target이 hard label 가정으로 fuzzy logic 일반화는 후속, 큰 DAG에서 학습 cost 증가 가능.

🚀 실용적 활용

산업 안전·감시 영상의 규칙 위반 탐지.
의료 영상의 semantic anomaly detection.
Neurosymbolic AI 학습 데이터 augmentation 일반화.