ABSTRACT. We recently developed a spiking neuron model that performs magnitude comparison and finger gnosis tasks using a common underlying neural system, explaining why performance on these tasks is associated in humans. Here, we explore the parameters in the model that may vary across individuals, generating predictions of error patterns across the two tasks. Furthermore, we also examine the neural representation of numbers in the magnitude comparison task. Surprisingly, we find that the model fits human performance only when the neural representations for each number are not related to each other. That is, the representation for TWO is no more similar to THREE than it is to NINE.

ABSTRACT. We explore the effects of parameters in our model of model-based reinforcement learning. In this model, spiking neurons are used to represent state-action pairs, learn state transition probabilities, and compute the resulting Q-values needed for action selection. All other aspects of model-based reinforcement learning are computed normally, without neurons. While some of these parameters have expected effects, such as increasing the learning rate and the number of neurons, we find that the model is surprisingly sensitive to variations in the distribution of neural tuning curves and the length of the time interval between state transitions.

ABSTRACT. Due to the correspondence between the striatal dopamine signal and prediction error signal utilized by model-free reinforcement learning methods, computational psychological research has found much success in modeling the basal ganglia as a biological implementation of a reinforcement learning mechanism. A large majority of these modeling efforts have focused on applying the tenets of reinforcement learning to the proposed functions of the basal ganglia, but few (if any) have attempted to apply crucial aspects of basal ganglia neurophysiology to reinforcement learning mechanisms. Here, we propose a basal ganglia-plausible model that explicitly utilizes two symmetric sets of actions (analogous to the basal ganglia's direct and indirect pathways), to simultaneously update value estimates of both available actions (i.e. chosen and not chosen) in the Probabilistic Stimulus Selection (PSS) task. We demonstrate that this proposed model architecture outperforms a standard reinforcement learning model of the PSS task by eliminating the standard model's bias towards estimation of the most valuable available actions, while granting improved resistance to noise in the internal selection process.

ABSTRACT. Building on earlier work extending Sigma’s mixed (symbols + probabilities) graphical band to inference in feedforward neural networks, two forms of neural network learning – target propagation and backpropagation – are introduced, bringing Sigma closer to a full neural-symbolic architecture. Adapting Sigma’s reinforcement learning (RL) capability to use backpropagation then yields a form of neural RL that is still combinable with probabilistic action modeling.

ABSTRACT. The explore-exploit dilemma occurs anytime we must choose between exploring unknown options for information and exploiting known resources for reward. Previous work suggests that people use two different strategies to solve the explore-exploit dilemma: directed exploration, driven by information seeking, and random exploration, driven by decision noise. Here, we show that these two strategies rely on different neural systems. Using transcranial magnetic stimulation to inhibit the right frontopolar cortex, we were able to selectively inhibit directed exploration while leaving random exploration intact. This suggests a causal role for right frontopolar cortex in directed, but not random, exploration and that directed and random exploration rely on (at least partially) dissociable neural systems.

ABSTRACT. Antisaccade performance in obsessive-compulsive disorder (OCD) is related to a dysfunctional network of brain structures including the (pre)frontal and posterior parietal cortices, basal ganglia, and superior colliculus. Previously recorded antisaccade performance of healthy and OCD subjects is re-analyzed to show greater variability in mean latency and variance of corrected antisaccades as well as in shape of antisaccade and corrected antisaccade latency distributions and increased error rates of OCD patients relative to healthy participants. Then a well-established neural accumulator model of antisaccade performance is employed to uncover the mechanisms giving rise to these observed OCD deficits. The model shows: i) increased variability in latency distributions of OCD patients is due to a more noisy accumulation of information by both correct and erroneous decision signals; (ii) OCD patients are almost as confident about their decisions as healthy controls; ii) competition via local lateral inhibition between the correct and erroneous decision processes, and not a third top-down STOP signal of the erroneous response, accounts for both the antisaccade performance of healthy controls and OCD patients.

ABSTRACT. We present an extension of a schema-based architecture for action selection, where competition between schemas is resolved using a variation of a neuroanatomically detailed model of the basal ganglia. The extended model implements distinct learning mechanisms for cortical schemas and for units within the basal ganglia. We demonstrate the functionality of the proposed mechanisms by applying the model to two classic neuropsychological tasks, the Wisconsin Card Sorting Task (WCST) and the Probabilistic Reversal Learning Task (PRLT). We discuss how the model captures existing behavioural data in neurologically healthy subjects and PD patients and how to overcome its shortcomings.

ABSTRACT. The Turing-inspired Meta-morphogenesis project begun in 2011 was partly motivated by deep gaps in our understanding of mathematical cognition and other aspects of human and non-human intelligence and our inability to model them. The project attempts to identify previously unnoticed evolutionary transitions in biological information processing related to gaps in our current understanding of cognition. Analysis of such transitions may also shed light on gaps in current AI. This is very different from attempts to study human mathematical cognition directly, e.g. via observation, experiment, neural imaging, etc. Fashionable ideas about ``embodied cognition'', ``enactivism'', and ``situated cognition'', focus on shallow products of evolution, ignoring pressures to evolve increasingly {\em dis}embodied forms of cognition to meet increasingly complex and varied challenges produced by articulated physical forms, multiple sensory capabilities, geographical and temporal spread of important information and other resources, and ``other-related meta-cognition'' concerning mental states, processes and capabilities of other individuals. Computers are normally thought of as good at mathematics: they perform logical, arithmetical and statistical calculations and manipulate formulas, at enormous speeds, but still lack abilities in humans and other animals to perceive and understand geometrical and topological possibilities and constraints that (a) are required for perception and use of affordances, and (b) play roles in mathematical, and proto-mathematical, discoveries made by ancient mathematicians, human toddlers and other intelligent animals. Neurally inspired, statistics-based (e.g.``deep learning'') models cannot explain recognition and understanding of mathematical {\em necessity} or {\em impossibility}. A partial (neo-Kantian) analysis of types of evolved biological information processing capability still missing from our models may inspire new kinds of research helping to fill the gaps. Had Turing lived long enough to develop his ideas on morphogenesis, he might have done this.

ABSTRACT. We describe a computational model of two central aspects of people’s probabilistic reasoning: descriptive probability estimation and inferential probability judgment. This model assumes that people’s reasoning follows standard frequentist probability theory, but is subject to random noise. This random noise has a regressive effect in probability estimation, moving probability estimates away from normative probabilities and towards the center of the probability scale. This regressive effect explains various reliable and systematic biases seen in people’s probability estimation. This random noise has an anti-regressive effect in inferential judgment, however. This model predicts that these contrary effects will tend to cancel out in tasks that involve both descriptive probability estimation and inferential probability judgment, leading to unbiased responses in those tasks. We test this model by applying it to one such task, described by Gallistel et al. (2014). Participants’ median responses in this task were unbiased, agreeing with normative probability theory over the full range of responses. Our model captures the pattern of unbiased responses in this task, while simultaneously explaining systematic biases away from normatively correct probabilities seen in other tasks.

ABSTRACT. Premises in conditional reasoning consist of an ``if'' statement (e.g., ``if I can catch the bus, I won't be late'') and a fact (e.g., I can catch the bus). Such types of simple inference have been studied empirically and formally for about a century. In the past five decades, several cognitive theories have been proposed to explain why humans deviate from predictions of conditional logic. In this article, we (i) describe existing theories, (ii) develop multinomial processing tree (MPT) models for these theories and systematically extend the theories with guessing subtrees to test the predictive power of cognitive models. The models are evaluated with G^2, Akaike’s (AIC) and Bayesian Information Criteria (BIC), and Fisher’s Information Approximation (FIA). Mental model theory with directionality for indicative conditionals while the independence model for counterfactuals provide the best fits to data from psychological studies.

ABSTRACT. Whether the mental representation of reasoning problems is spatial or visual (or mixed) in nature has been the subject of considerable debate for years. The visual impedance effect found in Knauff & Johnson-Laird (2002) has provided us with new insights into this question. The study found that the forming of excessive visual images induced by the premises can impede relational reasoning. This study aimed at investigating the factor of complexity on the visual impedance effect in two folds, namely number of term series (i.e., total number of premises plus the conclusion) and whether the entities in the premises are presented in a continuous manner (i.e., whether the subject of the argument is the same as the object of the previous argument). In line with previous studies, relational category, number of term series and successiveness were the main factors of the response time. Results of the parameter estimation by generalized estimating equation showed that visual relations, 5-term series and discontinuous problems were the only significant parameters. The results again suggested that irrelevant visual images can hinder reasoning processes, in addition to the complexity of the problem. We proposed a combined cognitive model of ACT-R and PRISM for the findings in this study.

ABSTRACT. This paper demonstrates how mental models and updates of mental models due to system changes can be modeled with the cognitive architecture ACT-R using explicit mechanisms. The mental model building and updating is modeled with a representation chunk and a control chunk. The representation chunk holds the strategy, the expected outcome and an evaluation mechanism of the strategy. The control chunk holds information over environmental conditions and the learning history. This modeling approach was developed and tested for smartphone application tasks and then implemented in a dynamic decision making task investigating strategy development with complex stimuli. The later task used different multi-feature auditory stimuli material. The modeling approach explained data of participants in the smartphone studies very well and met the trends found in the dynamic decision making task.

ABSTRACT. In the area of sequential choice, the ‘Secretary Problem’ has been a prominent paradigm within the study of optimal stopping for sequential search tasks. Most recent studies of the Secretary Problem present decision makers with the relative ranks of options. A recurring finding is that decision makers tend to end their search earlier than optimal decision strategies (e.g. Helversen, Wilke, Johnson, & Schmid, 2011; Seale & Rapoport, 1997, 2000). By revealing only relative ranks of options or items, issues of learning and incomplete knowledge are avoided; however, this leaves open the question of how sensible human decision makers are when they know more about the distribution of items. Rather than presenting merely ranks to decision makers, we presented numerical values drawn from three distinct distributions in which relatively high value items were scarce, evenly distributed, or abundant. We found that they selected their items earlier than they would if they utilized the optimal selection rule. More importantly, in contrast to the conclusion of Kahan, et al. (1967), we found the selection points of decision makers were sensitive to the underlying distribution. In contrast, the optimal strategy is totally based on quantile ranks regardless of the type of distributions.

ABSTRACT. Decisions from Experience (DFE) research involves a paradigm (called, sampling paradigm), where decision-makers search for information before making a final consequential choice. Although DFE research involving the sampling paradigm has focused on accounting for information search and final choices using computational cognitive models. However, this research has yet to investigate how computational models would account for final choices for participants with different sampling strategies during information search. In this paper, we perform an individual-differences analysis and test the ability of computational models to explain final choices of participants with different sampling strategies. More specifically, we take an Instance-Based Learning (IBL) model, which relies on recency and frequency processes, and we calibrate this model to final choices of participants exhibiting more-switching (piecewise) or less-switching (comprehensive) between options in different problems. Our results indicate more reliance on recency and frequency of information among participants exhibiting piecewise strategy compared to comprehensive strategy. Overall, the IBL model is able to account for piecewise strategy better compared to comprehensive strategy. We highlight the implications of our results for DFE research involving information search before consequential decisions.

ABSTRACT. A novel quantum cognition model is proposed to address the conjunction and disjunction fallacies. This model represents each concept as a separate qubit and the measurement process as a weak measurement. In order to evaluate our model we conducted an on-line survey questionnaire that addressed several conjunction and disjunction fallacies scenarios that included four different concepts. The novel model enables us to calculate a quantitative measure of quantum entanglement for each participant and each question. We show that irrational judgment is represented by an entangled quantum state, whereas a separable state represents a rational judgment, in both conjunction and disjunction fallacies. Following individual participants' quantum cognitive representation throughout the questionnaire shows their entanglement dynamics. These results suggest a deeper connection between the quantum representation of cognitive concepts and the ensuing irrational judgments, namely, that quantum entanglement between mental states are correlated to irrational behavior regarding these concepts.