Mental Representation

People’s mental representations of complex stimuli, such as facial expressions, are difficult to elicit. To address this challenge, methods such as Markov Chain Monte Carlo with People (MCMCP) integrate human decision-making into computer-based sampling algorithms. However, MCMCP suffers from slow convergence due to the high-dimensional sample space and inefficient navigation, making it impractical for recovering representations of individuals. Here, we extend MCMCP by combining it with an adapted Variational Auto-Encoder (VAE) that addresses the problem of slow convergence in two ways: 1) it usefully represents the facial expression of images in only three latent dimensions, and 2) it acts as a gatekeeper, using its own domain knowledge to quickly reject less useful trials to save participant effort. We tested this approach in a new experiment (N=90) on facial affect comparing baseline MCMCP (in a 157-dimensional PCA space) against both MCMCP in the 3-dimensional VAE latent space and MCMCP with a gatekeeper (MCMCPG) in the same VAE latent space. MCMCPG converged substantially faster than the other methods, generally in less than 200 trials, and the average recovered face was judged to be more representative than those from PCA and those from past work. Further analyses also revealed the extent of individual differences in facial affect representations, which could partially account for individual differences in decision-making. Our study demonstrates the potential of MCMCPG for investigating human representations at the individual level. It also provides a promising framework for linking machine intelligence with psychological experiments to enhance our understanding of human cognition.

This is an in-person presentation on July 21, 2024 (15:20 ~ 15:40 CEST).

In our daily lives, encountering exact replicas is a rarity due to the ubiquitous presence of variability. As humans, we possess the remarkable ability to extrapolate from past experiences and adapt to diverse contexts, even when they are not identical. However, the mechanisms underlying this transfer process remain largely enigmatic, partly because current investigations often assume that variability is confined solely to the external features of physical inputs. This overlooks the intricate dynamics involved in how organisms interpret and process the physical world. To address this gap, we developed a computational model that amalgamates error-driven learning and similarity-based generalization processes. By integrating these processes, we endeavor to unravel the underlying mechanisms that govern our ability to generalize across diverse contexts. Presently, the three studies we have conducted with the computational model show that learning and mental representation are indeed important sources of adaptive behavior, that during the generalization process, mental representation is often probabilistic, inferential, and dynamic rather than static, and that mental representation during generalization process has specific temporal dynamics which are context-dependent.

This is an in-person presentation on July 21, 2024 (15:40 ~ 16:00 CEST).

Natural environments are feature-rich and only a subset of these features is considered to predict action-outcome associations. To enable accurate action-outcome predictions a decision maker is faced with a challenge, namely that only a portion of the information in the environment is predictive of a desired outcome. Here, we highlight across several studies the tendency of individuals to assign credit to outcome-irrelevant task representations. We demonstrate that value is assigned to these representations in a model-free and state-independent manner. We further show the association between these low-level value associations and a more sophisticated model-based system and propose how model-free representations might be regulated according to a model of the environment. Finally, we suggest that a deficit in the regulation of outcome-irrelevant model-free associations might lead to behavioral abnormalities such as compulsive behavior. These findings call for a revision of current reinforcement-learning models which are largely based on state-dependent and outcome-relevant learning.

This is an in-person presentation on July 21, 2024 (16:00 ~ 16:20 CEST).