Beliefs & Selective Attention

Human beliefs are often assumed to be irrational and belief updating is viewed as suboptimal. For example, Ward Edwards (e.g., 1968/1982) famously found that human probability judgments are conservative. However, Edwards also observed that – under certain conditions – judgments often fail to show this conservative bias. Sommer, Musolino, and Hemmer (2023) proposed a theory of belief which distinguishes belief updating from evidence evaluation. Updating is argued to be approximately Bayesian and inaccessible to consciousness. In contrast, evidence evaluation processes are consciously accessible and may be the true cause of apparently non-Bayesian beliefs, including conservatism. We suggest that the presence of conservatism depends on whether a judgmental task requires updating alone or whether evidence evaluation processes are also necessary. Here we test this hypothesis by presenting participants with two versions of judgment problems adapted from Edwards’ work, where participants judge the probability of marbles in urns. We manipulate whether the marbles are shown sequentially (the updating alone condition) or all at once (the evidence evaluation condition). We find a difference in optimality between the two conditions, as well as substantial individual differences. Notably, we find that more participants are optimal in the sequential condition. We implement a mixture model to capture individual judgment strategies. We assume that participants are a mixture of five strategies: Normative updating, conservatism, “stick with the prior”, frequentist, or random guessing. We discuss the implications of our results in the context of processes involved in belief.

This is an in-person presentation on July 22, 2024 (10:00 ~ 10:20 CEST).

Categorization often involves the use of episodic memory (Nosofsky, 1986; Turner, 2019), and category learning is thought to involve the hippocampus (Mack, Love, & Preston, 2016). In the hippocampus, memory retrieval can be modeled as pattern completion in an attractor network, a recurrent network that can reconstruct previously-observed patterns through complex, dynamic interactions between nodes (e.g., Rolls, 2007). Attractor networks can also be used to model hippocampal learning of sequential dependencies, such as transitive inference and sequence learning, as well as other forms of learning including reinforcement learning (Whittington et al., 2020). In this work, we show that many models of category learning have a natural completion as a two-layer, discrete-time attractor network (a Dense Associative Memory or Modified Hopfield Network; Krotov & Hopfield, 2021; Ramsauer, 2021). We call this meta-model AuToassociative and HEteroassociative Neural Attention (ATHENA). From this perspective, a wide range of mechanisms proposed to explain categorization, as well as common techniques in machine learning, can be implemented in the same architecture through the use of different input representations, forms of memory competition, and learning mechanisms. Furthermore, when augmented with idealized, brain-inspired connections, models in the ATHENA family exhibit flexible sequence learning. We show through simulations that this ability can be used to achieve accelerated replay of observed sequences, multi-step transitive inferences, and one-shot reinforcement learning. Together, this suggests that category learning models can be implemented as special cases of a domain-general hippocampal learning mechanism, providing a link between cognitive models and complex neural architectures.

This is an in-person presentation on July 22, 2024 (10:20 ~ 10:40 CEST).

The queueing model of visual search (Li, Schlather, & Erdfelder, 2023) was developed for visual search in which attentive processing is necessary for the final decision, as it aims to explain the so-called attentional bottleneck, that is, the allocation of attentional resources when transiting from preattentive processing to the more resource-intensive attentive processing. Although the model’s explanatory power was supported by the good model fit to empirical conjunction search and spatial configuration search data on a distributional level, whether it can account for feature search is still open. Feature search is important to understand preattentive proccessing and empirical data of feature search provide incremental information for the examination of the model assumptions. In this presentation, I will explain from a technical perspective why the adaptation of the queueing model of visual search to feature search is not a simple reduction but rather an extension. Then I will introduce different approaches of adaptation, compare their advantages and disadvantages using simulation. Finally, fitting the adapted model to empirical data of feature search will be discussed and compared to the result of fitting to conjunction search data based on the same visual material.

This is an in-person presentation on July 22, 2024 (10:40 ~ 11:00 CEST).

The rule-based and information integration tasks have been a staple across a myriad of experiments in comparative psychology as a means to test for the presence of selective attention through the relative differences in the speed of learning. Specifically, rule-based tasks are generally learned faster relative to information integration tasks for learners who possess selective attention, whereas the two tasks are learned equally quickly for learners who lack selective attention. Although Smith et al (2012) documented RB vs. II performance across four species, less is known about species such as rats who have reportedly expressed selective attention. In addition, we present a new experiment involving a switch from one subtask to another. For example, if learners first perform an RB task in which the rule is associated with Dimension 1, then after the switch learners perform another RB task in which the rule is now associated with Dimension 2. This unique manipulation allows us to detect the presence of selective attention by examining how knowledge is transferred from one phase of the experiment to the next. We report results from this experiment on four species: pigeons, rats, rats with prefrontal cortex lesions, and humans.

This is an in-person presentation on July 22, 2024 (11:00 ~ 11:20 CEST).