Nadav Amir

ABSTRACT A fundamental problem facing any learning organism is how to determine which aspects of its experience to attend to and which to ignore when guiding its goal-oriented actions. This problem, known as relevance realization, is key to understanding how cognitive agents, biological or artificial, learn purposeful behaviours through experience-based interactions with their environment. However, a comprehensive theoretical understanding of how organisms solve the relevance realisation problem is currently lacking. Here, we approach this challenge by proposing a synthesis between a contemporary computational framework of goal-oriented learning in cognitive agents and a canonical Buddhist account of concept formation, specifically as developed by the Indian philosopher Dharmakīrti (c. 6-7th century C.E.). Our approach posits that cognitive agents form goal-dependent conceptual representations of the world that reflect subjective preferences about properties of simulated sensorimotor experience streams. By drawing on Dharmakīrti’s ideas to inform computational models of goal-directed learning, our approach bears on current debates regarding the prerequisites for and implications of artificially emulating human-like cognition.

Recent developments in biology and cognitive science have called into question the fecundity of the notion of an ‘agent’—conceived as a unitary and enduring locus of representation, evaluation, and autonomous action—for explaining the phenomenology of purposeful behavior across diverse substrates and spatio-temporal scales. A coherent account of goal-directed behavior in cognitive systems, without strong commitments to the aforesaid notion of ‘agent’, is therefore desirable, particularly in light of increasingly blurred boundaries between biological and artificial agentic systems. Here we propose that Buddhist philosophy, with its central doctrine of no-self, offers a compelling foundation for substrate-independent accounts of agency by replacing autonomous and enduring agents with stable yet malleable goals. Furthermore, according to Buddhist views, goals systematically bias cognitive representations, resulting in distorted self-models which are the root cause of suffering. We therefore posit that Buddhist-based models of agency in cognitive systems can organically integrate the epistemic, phenomenological, and ethical aspects of purposeful behavior under a unified theoretical framework. We furthermore argue that a cross-disciplinary account of agency, drawing on Buddhist notions, can help address emerging risks posed by artificial and hybrid agentic technologies.

A fundamental problem facing any learning organism is how to determine which aspects of its experience to attend to and which to ignore when guiding its goal-oriented actions. This problem, known as relevance realization, is key to understanding how cognitive agents, biological or artificial, learn purposeful behaviours through experience-based interactions with their environment. However, a comprehensive theoretical understanding of how organisms solve the relevance realisation problem is currently lacking. Here, we approach this challenge by proposing a synthesis between a contemporary computational framework of goal-oriented learning in cognitive agents and a canonical Buddhist account of concept formation, specifically as developed by the Indian philosopher Dharmakīrti (c. 6-7th century C.E.). Our approach posits that cognitive agents form goal-dependent conceptual representations of the world that reflect subjective preferences about properties of simulated sensorimotor experience streams. By drawing on Dharmakīrti’s ideas to inform computational models of goal-directed learning, our approach bears on current debates regarding the prerequisites for and implications of artificially emulating human-like cognition.