CAVAA vision

A Step Towards AI Awareness

In real-world implementations artificial agents fall short in terms of spatial and temporal awareness, awareness about self and others, and decision-making abilities. Through CAVAA, the consortium plans to generate artificial awareness by interfacing a cognitive architecture agents with the propensity of reasoning, decision-making, or revisiting past experiences and reflecting upon what was right or wrong.

The international consortium will follow a structured 4-year plan to achieve the step toward AI awareness. CAVAA will elaborate an underlying theoretical framework of awareness and express it as cognitive architecture that aims to provide an important contribution towards the explanation of awareness in a biological system and to engineer it in technological agents that collect and store experiences while interacting with the exterior world. This will be achieved by building upon a well-recognized, tested Distributed Adaptive Control (DAC) architecture. A new virtualization layer will be added to represent potential states of the world and the self, entirely decoupled from immediate sensory states, compressed into a single unified awareness state, and its top-down control over memory, valuation, and creativity.

Can AI have awareness? Will machines be able to evaluate the external context and implementing appropriate action plans? CAVAA can be a revelation and help solve age-old questions about AI awareness.

CAVAA challenges

in philosophy of mind, brain and cognitive research, machine learning, and robotics

Following Thomas Nagel, conscious agents have an awareness of what it is like to be that agent. This first- person definition precludes a third person scientific pursuit leading to the, so called, hard problem or explanatory gap, with its pragmatic solution of distinguishing phenomenal from access consciousness. Conscious awareness can further be characterized in terms of its level, from coma to alertness, and content or quale pertaining to the distinction between the outside world and the self and the level of abstraction.

In the face of these definitional challenges theories of consciousness have been emphasizing different non-exclusive aspects such as grounding in the self and sensorimotor contingencies, complexity, information access, prediction, attention, or meta-representations. However, none of these theories give any hypothesis on what the function of consciousness could be, let alone its role in future technology and artificial intelligence systems.

The Counterfactual Assessment and Valuation for Awareness Architecture (CAVAA) project dissents from this position. CAVAA proposes that awareness has a specific function in the control of adaptive behavior which emerged during the Cambrian explosion: the ability to survive in a world governed by hidden states, especially those pertaining to other agents. Indeed, CAVAA proposes that awareness allows agents to decouple from immediate sensory states and deal with the “invisible”, ranging from behavioral policies in unexplored environments and unobservable aspects of tasks, to the complexities of social interaction that depend on the internal states of agents (e.g. intentions, knowledge, and emotions), and the moral norms that guide collective interaction action. Awareness thus reflects a virtual world which is a hybrid of perceptual evidence, memory states, and inferred “unobservables”, extended in space and time and which rests on five core processes: the ability to virtualize task spaces, to merge “real” and virtual elements into these internal models, to run parallel future-oriented simulations of possible world-self states, to collapse these into a single conscious scene which defines the content of awareness, and to use awareness to bias valuation and memory consolidation.

In realizing this goal, the CAVAA consortium will build on our growing understanding of the biological basis of awareness and its role in constructing the virtualized world models in which mental life occurs. The CAVAA project starts with an explicit theory on awareness focussing on virtualisation, simulation, compression, and valuation, which is realised in an awareness architecture, thus facilitating synthetic awareness in AI systems.

As a result, CAVAA delivers concrete hypotheses on what users of such an aware system can expect in terms of trust, transparency, and empathy through shared ontologies, intention complementarity through other-like-self mirroring and extended self-other simulation, including counterfactuals and insight. This formulation also facilitates a generalisation of awareness beyond anthropocentric concepts because the virtualisation, simulation, compression, and feedback cycle as defined by CAVAA, can occur in various memory-based systems from ant colonies to culture and AI systems, extending the elusive notion of extended minds to that of extended awareness. CAVAA thus fully accommodates the notion of variability in the processes and quality of awareness across agents and collectives, facilitating inter-subjective coherence because it can identify potential incompatibilities between awarenesses and provide translation channels, e.g. deliver a meta-cognitive understanding of mismatches in ontologies, capacity limitations, or awareness biasing of perception, memory and valuation. The CAVAA project explicitly validates its proposition by its concrete commitments to operationalising the concepts underlying its theoretical framework.