Representational Alignment and Aging

Motivations, goals, and rewards are the building blocks of RL, but where do they come from? In traditional RL models we typically ignore these questions, but in the study of human cognition they are front and center. For example, Socioemotional Selectivity Theory (SST) describes how people select and prioritize goals and how their motivations change - especially as they age and their perceived future time horizons are shortened. According to SST, as they age humans increasingly prioritize social, emotional, and physical well-being over optimizing behavior in the pursuit of extrinsic rewards (e.g., money, prizes). Yet, how to align these socioemotional functions into an integrative theory of learning over the lifespan is unknown. This intuitive understanding that an agent’s perceived time horizon influences the evolution of goal selection (either in biological or artificial age) has yet to be formalized computationally in models of reinforcement learning and decision-making. This representational alignment is critical for moving towards a holistic understanding of the mechanisms that underlie how motivational and affective functions are represented in both natural and artificial agents across the lifespan. 

This workshop aims to bridge the gap between the representation of socioemotional functions in natural and artificial intelligence by bringing together experts across diverse interdisciplinary perspectives (e.g., computational and/or affective neuroscience, computer science). First, we will introduce several key challenges of representational alignment of socioemotional function between human and artificial agents. We also highlight several avenues that may be fruitful for moving the needle toward building a clearer understanding of how agents select and prioritize goals to optimize their decision-making across the lifespan. Our second speaker will discuss the central theoretical assumptions of Socioemotional Selectivity Theory, both in terms of its strengths and limitations in explaining motivated human decision-making and key features of SST theory most relevant to artificial intelligence. Our third set of speakers will present current findings on model-based and experimental approaches for characterizing motivational/affective changes in humans across the adult lifespan. Our fourth set of speakers will highlight recent advances in autonomous learning and large language models and address the need to build machines and algorithms that learn conceptual representations informed by human cognitive and affective processes. Our fifth set of speakers will highlight the need to translate such computational models to provide novel insight into understanding how motivational, affective, and cognitive functions are altered in healthy and pathological aging (e.g., Alzheimer’s Disease and Related Dementias). Finally, we will conclude with an interactive panel discussion (with audience participation) to brainstorm ways in which researchers across these diverse fields can bridge the gap of representational alignment – by incorporating insights of socioemotional function into RL as well as establishing novel computational frameworks for testing normative assumptions of Socioemotional Selectivity Theory in natural and artificial agents.