About Brainglot
Presentation
History
Objetives
Summary
Brainglot personnel
 
 
Activities
Research Areas and Projects
PhD
Publications
Media
Consolider CogNeuro Seminar Series
Consolider Meetings and Workshops
Technical and Administrative Support Services
Research Reports
 
 
Consolider Groups
Speech Perception, Production and Bilingualism
Cognitive Neuroscience of Auditory Perception and Attention
Group of Attention, Action and Perception
Computational and Theoretical Neuroscience
Neuropsychology and Functional Neuroimaging
Grammar and Bilingualism
 
 
 

Psychophysical and theoretical studies on flexibility (changes of mind) in decision-making

Project coordinator: Gustavo Deco.

Project members: Larissa Albantakis, Francesca Branzi, Clara Martín, Albert Costa.                                                                                                    
       

GOALS OF THE PROJECT 

Traditionally, the decision process is regarded as a decision variable evolving in time until a criterion value is reached, which terminates the decision process. Correspondingly, perceptual decision-making tasks usually measure reaction time and end when the subject reports his/her choice. In real life, however, first choices are not necessarily final: choices can be re-evaluated and this may lead to a change of mind. The present project focuses on this more complex aspect of decision-making. Some of the central questions are: What happens in our brains during changes of mind? Does the brain continue to evaluate available evidence even after an initial decision is taken? What criterion is applied when deciding to change? How is the number of changes affected by the number of decision alternatives? To shed light on these questions we use a combined approach of computational modelling coupled with psychophysical decision-making experiments.  

SUBPROJECT 1: Changes of mind in an attractor model of decision-making.   An experimental paradigm that matches the abovementioned notion of the decision-process is the random-dot motion (RDM) task. Here, subjects have to decide on the net motion of a cloud of (randomly) moving dots and typically report their choice with a saccade to one of two possible targets. By altering the way subjects indicate their choice from ballistic saccades to continuous hand movements, changes of mind can be observed directly by recording the trajectory towards the target. Our research showed that changing improved the overall accuracy, although this depended on task difficulty: most correcting changes were observed at intermediate levels, while erroneous changes increased monotonically with difficulty. These experimental findings pose a challenge for a class of models that implement decision-making by diffusion in a nonlinear landscape of stable fixed points, which act as decision-attractors. Over the past decade attractor networks have successfully modelled psychophysical and neurophysiological data in various kinds of decision-making tasks. In particular, one biophysically realistic implementation reproduces and accounts for neurophysiological recordings from LIP neurons during the random-dot motion task. A change of mind, however, is rather counterintuitive to the stable character of attractors, as once a decision-attractor is reached this state will persist, except for high levels of noise or perturbations. In this subproject we explore the capacity of attractor models to account for changes of mind. Specifically, we aim to elucidate the neural mechanisms involved in choice re-evaluation and to offer predictions on neural firing rates during changes of mind. The choice alternatives in the biophysically realistic attractor model are implemented by two neural populations of excitatory neurons that selectively respond to motion towards one of the two choice targets. The decision process corresponds to the transition from a symmetric state, where both selective populations fire with about equal rates, to a decision state where they compete with each other in a winner-take-all manner, resulting in one pool firing at higher rates (winner) and the other at lower rates (loser) (Figure 1 A). Under the assumption that the brain continues to evaluate available evidence even after making a first decision, we found that attractor models are capable of capturing the essential aspects of changes of mind during the dynamic transitions to their steady states (decision-attractors) (Figure 1 A,B). Moreover, the frequency of changes in the attractor network is positively related to an emphasis on reaction speed versus accuracy. In the model, the speed-accuracy trade-off can be adjusted either by adapting the decision threshold, or by increasing or decreasing the inputs to both selective populations. Both of these mechanisms (i.e. a low decision threshold and high inputs) were found to be essential in order to fit the experimental data. Technically, high incoming activity shifts the working point of the system into a dynamic regime close to a bifurcation between the region of categorical decision-making and a multi-stable region. The proximity of this attractor impedes the decision process and thereby facilitates changes of mind. In this context, we found evidence for the physiological relevance of a so-far unregarded bifurcation in the binary attractor model, thereby confirming the general accordance of attractor networks with neural processes.

SUBPROJECT 2: Psychophysical experiments on changes of mind in relation to the number of choice alternatives.   Changes of mind have so far only been investigated during binary perceptual decision-tasks, the simplest case to study. In a series of psychophysical experiments we intend to elucidate more complex aspects of decision-making and changes of mind, starting with the relationship of change-frequency to the number of choice alternatives. To that end we designed a perceptual task in which human subjects have to decide on the most frequent colour in a square matrix of coloured dots and then indicate their choice by moving a computer mouse to the respective colour target located at the corners of the computer screen. The task difficulty is set by the dot-percentage ratio of the different colours. In preliminary experiments with two choice alternatives, and thus two colours to decide on, we were able to replicate qualitatively the difficulty-dependent frequency of changes from the binary RDM task (see subproject 1), but using a much simpler experimental setup (Figure 1 C). We also tested subjects on the same task with four colours (Figure 1 D). Difficulty was determined by the percentage of dots of the dominant colour, while all other colours had the same number of dots (e.g. 70% red and 10% dots of each of the remaining three colours). Expanding the paradigm to four possible choice alternatives increased the difficulty of the task substantially and apparently added stimulus-independent sources of confusion to the task, which also caused changes of mind. For example, and surprisingly, participants changed their mind occasionally even when all the dots had the same colour, thereby suggesting confusion about target location. By isolating the different sources of changes of mind, for example, by intensively training participants to avoid confusion about target location, we hope to gain further insights in the process of choice re-evaluation.  

           

 

 

 

Figure 1.  Changes of mind – theory and experiments. (A,B) Simulated LIP firing rates during changes of mind. (A) Single trial. Light and dark blue: neural populations each encoding one of two choice alternatives, black: nonselective neurons, grey: inhibitory neural population. (B) Mean of all trials with changes of mind aligned to first threshold crossing. (C,D) Hand-movement traces recorded during psychophysical experiment on changes of mind with two (C, 100 trials) and four (D, 200 trials) choice alternatives.

 

OTHER SUBPROJECTS
  1. Four-colour decision task with different divisions of non-dominant colours (e.g. 40%, 20%, 20%, 20% compared to  40%, 30%, 20%, 10% of dots)
  2. Computational model for changes of mind during a multiple-choice task.