Periodic Reporting for period 2 - CAVAA (Counterfactual Assessment and Valuation for Awareness Architecture)
Período documentado: 2023-10-01 hasta 2025-03-31
The Counterfactual Assessment and Valuation for Awareness Architecture (CAVAA) project centers on the concept that awareness is vital for survival in a world of unseen elements and internal dynamics of agents and moral norms. This awareness combines perceptual data, memory, and inferred elements to form an internal virtual world. CAVAA aims to develop an integrated computational architecture to understand and engineer awareness in both biological and technological systems. The project will focus on perception, memory, virtualization, simulation, and integration, and apply these in robots and artificial agents across various use-cases like robot foraging, social robotics, and computer game benchmarks. These scenarios will test trade-offs such as efficiency versus robustness and gauge user acceptance.
During Period 2, major progress was made in developing and integrating the CAVAA architecture across multiple cognitive layers. In WP1, three biologically inspired modules were implemented: (1) the Reactive Layer uses a neural mass allostatic model to replicate hypothalamic dynamics and support self-regulatory behavior; (2) the Adaptive Layer introduces a Motivational Hippocampal Autoencoder, allowing artificial agents to build self-referential cognitive maps; and (3) the Contextual Layer includes Sequential Episodic Control, a hippocampal-inspired memory algorithm that enhances sample-efficient reinforcement learning. These modules have been benchmarked and integrated into the full CAVAA architecture. The Virtualization Model, now fully operational, explains a wide range of hippocampal replay phenomena and has been successfully embedded into the system. In WP2, novel models at the interface of cognitive science and deep learning were developed, including models of attention, goal-directed decision-making, and learning under virtualization. These contribute to both model-based and model-free learning, leveraging cognitive mechanisms such as visual attention and memory. Progress in explainable AI and self-supervised vision transformers resulted in two peer-reviewed publications. In WP3, key components were developed to support physical embodiment and interaction with real or simulated environments. A sensory acquisition interface was implemented to feed environmental data into the architecture. A motor command interface was built to translate the architecture’s outputs into discrete or continuous motion commands. The Motivational Hippocampal Autoencoder was trained in a warehouse environment. WP4 delivered new standardized measures of awareness for biological and artificial systems, aligned with human subjective evaluations. WP5 enabled rich interdisciplinary collaboration between engineers, neuroscientists, ethicists, and philosophers. This led to several notable publications on technical, ethical, and conceptual aspects of AI awareness, including: a new classification of strong vs. weak AI alignment with human values; conceptual clarifications on artificial consciousness; and innovative work on value alignment in moral dilemmas using large language models and probabilistic reinforcement learning. A two-day workshop on AI awareness and privacy was held at the University of Oxford in May 2025, bringing together academic and industry stakeholders, to advance the scientific and societal understanding of artificial awareness, particularly in relation to agency, responsibility, and privacy.
In Period 2, the CAVAA project significantly extended its cognitive architecture, advancing beyond existing models of decision-making and awareness in artificial systems. The integration of biologically inspired modules—such as the neural mass allostatic model, the motivational hippocampal autoencoder, and the Sequential Episodic Control represents a novel hierarchical framework that brings together reactive, adaptive, and contextual processing. The operational Virtualization Model provides a unique capability for mental simulation and has been validated against neuroscientific data on hippocampal replay. In addition, the development of new attention and decision-making models tightly coupled with deep learning significantly improves agent perception. These models offer a unified framework for model-based and model-free learning through interpretable mechanisms such as visual attention and cognitive maps. The project also made major theoretical and ethical contributions. In particular, newly defined metrics for measuring awareness in artificial agents, alongside conceptual advances in AI alignment and artificial consciousness, provide foundational tools that the field previously lacked. These are now supported by several high-profile publications and community-building efforts, such as the 2025 Oxford workshop on AI awareness and privacy. Together, these developments position CAVAA at the forefront of research into conscious, explainable, and ethically grounded AI systems.