Combing Perception with Structured Knowledge for Rich Causal Reasoning in a Computational Cognitive Architecture
Abstract
The ARCADIA computational cognitive architecture was extended with significant new capabilities to visually extract, encode, and use relational information present in videos, along with having been endowed with an episodic memory capable of storing hierarchical representations of temporally ordered events and episodes. ARCADIAs existing capabilities for visually verifying causal relations was extended to handle the tricker case of omitted events being recognized as causes. These capabilities were brought together and married with a simplified approach to counterfactual reasoning in which ARCADIA was able to deliberate about what would have happened in a video sequence had an omitted event (or more than one omitted event) actually happened. Secondly, a psychologically plausible model of the blame attribution process was implemented in ARCADIA allowing the system to visually inspect text strings that described interactions between agents where moral violations were described and to output initial and subsequent updated attributions to blame as more text describing the context of the violation was provided. The model was tested on published human data and matched using the same stimuli with a good fit being obtained. Finally, several human subject studies were done looking to investigate how humans represent and reason about causes interact with norms and other situational constraints. In summary, the results of these three project thrusts provide some initial evidence that it may be possible to have an autonomous system be able to perceptually parse up complex social interactions between agents in situations where norms are violated and to initially apportion and update blame judgments accordingly. It seems quite reasonable to suppose that they would make a difference to trust and other dimensions of human-machine ormachine-machine teaming.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 06, 2023
- Accession Number
- AD1212712
Entities
People
- Paul Bello
Organizations
- United States Naval Research Laboratory