Periodic Reporting for period 3 - INFOSAMPLE (Information Sampling in Multiattribute Choice)
Okres sprawozdawczy: 2021-09-01 do 2023-02-28
Understanding how humans make decisions is a central theme in the behavioural sciences, with several interdisciplinary ramifications. Recent studies in the fields of psychology and neuroscience have shed light on the neural and computational mechanisms via which humans and other animals make sensorimotor transformations in simple laboratory tasks. However, little is known about the mechanisms underlying complex multiattribute decisions. For instance, what mechanisms does the brain employ when we choose between a small flat with a short commute or a large house with a long commute?
Existing computational mechanisms underlying sensorimotor integration do not apply to multiattribute decisions. This is due to the fact that human choices in multiattribute settings are at odds with rational explanations of behaviour. To illustrate, an initial preference for the small flat in the centre can be reversed for the large house in the outskirts, soon after a medium-sized (and thus inferior) house in the outskirts becomes available. Contextual preference reversal of this type currently lack a single and neurophysiologically-grounded theoretical explanation. Elucidating the mechanisms of multiattribute choice requires understanding how humans sample and accumulate information when presented with complex alternatives, characterised in more than one attribute. I am addressing this question in a data-driven fashion and by harnessing tools from sensory neuroscience. Using magnetoencephalography (MEG) and perceptual analogues of multiattribute decisions, we will simultaneously track the locus of attention and the tendency to choose one alternative over the other, during the entire time-course of a single decision. This approach will enable us to unravel the computational and neural mechanisms that guide attention towards different aspects of a multiattribute choice problem.
This project will yield a neurophysiologically detailed theory of multiattribute choice and the emerging framework will be useful to policy makers and practitioners, interested in a descriptively enriched model of choice; and to clinicians aiming to understand how information sampling goes awry in neuropsychiatric disorders.
The first objective of the project is to understand the computational principles that guide voluntary information sampling. The second objective is to understand the neural mechanisms that orchestrate information sampling during multiattribute decisions.The third objective of the project is to develop a biophysically constrained algorithmic model of multiattribute choice, that will be used to fit well-known behavioural patterns (e.g. preference reversals) as well as to simulate choice behaviour under altered neurotransmitter profiles, that are hallmarks of various neuropsychiatric disorders.
Existing computational mechanisms underlying sensorimotor integration do not apply to multiattribute decisions. This is due to the fact that human choices in multiattribute settings are at odds with rational explanations of behaviour. To illustrate, an initial preference for the small flat in the centre can be reversed for the large house in the outskirts, soon after a medium-sized (and thus inferior) house in the outskirts becomes available. Contextual preference reversal of this type currently lack a single and neurophysiologically-grounded theoretical explanation. Elucidating the mechanisms of multiattribute choice requires understanding how humans sample and accumulate information when presented with complex alternatives, characterised in more than one attribute. I am addressing this question in a data-driven fashion and by harnessing tools from sensory neuroscience. Using magnetoencephalography (MEG) and perceptual analogues of multiattribute decisions, we will simultaneously track the locus of attention and the tendency to choose one alternative over the other, during the entire time-course of a single decision. This approach will enable us to unravel the computational and neural mechanisms that guide attention towards different aspects of a multiattribute choice problem.
This project will yield a neurophysiologically detailed theory of multiattribute choice and the emerging framework will be useful to policy makers and practitioners, interested in a descriptively enriched model of choice; and to clinicians aiming to understand how information sampling goes awry in neuropsychiatric disorders.
The first objective of the project is to understand the computational principles that guide voluntary information sampling. The second objective is to understand the neural mechanisms that orchestrate information sampling during multiattribute decisions.The third objective of the project is to develop a biophysically constrained algorithmic model of multiattribute choice, that will be used to fit well-known behavioural patterns (e.g. preference reversals) as well as to simulate choice behaviour under altered neurotransmitter profiles, that are hallmarks of various neuropsychiatric disorders.
At the beginning of the project strong emphasis was placed on developing experimental and analytical tools that, using MEG, would enable the temporally precise tracking of information sampling and accumulation during multi-alternative, multiattribute decision-making tasks. Towards this end, my team and I tested the accuracy of various decoding techniques (inverted encoding models, representational similarity analysis) in control tasks that a) involved overt and covert instructed shifts of attention, b) instructed motor actions using one out of four (two hands, two feet) effectors. At the same time we tested the potential of rapid frequencies (above 70 HZ) in generating steady-state-visually evoked potentials (SSVEP’s) that could track the locus of attention in instructed attention tasks.
The explorations and technical developments above were exploited in the first MEG experiment (N=20, 2 5-hour sessions per participant). In this experiment we presented to participants 3 Gabor patches against grey background, whose contrast levels and orientations varied, and across different trials asked them to select either the “highest” or the “lowest” contrast. Thus, unlike multiattribute decisions in which both dimensions are relevant, in this paradigm orientation was an irrelevant dimension. At the behavioural level we observed the so-called “distractor” effect, previously occurred in unidimensional decisions (according to which the higher the value of the worst alternative is the lower the probability of telling apart the two best alternatives is). At the same time, and despite the fact that orientation was in this experiment irrelevant, we observed well-known multiattribute contextual preference reversal effects (e.g. the attraction effect). These findings indicate that the perceptual (dis-) similarity among alternatives, even on the basis of irrelevant features, biases the way information is sampled and processed. At the neural level, we are currently focusing on how the decoded sensory representation associated with the three alternatives changes in the course of a trial a) as a function of the task framing (“highest” vs. “lowest”), b) as a function of the value of the worst alternative.
The above-mentioned MEG experiment was coupled with a similar behavioural experiment (N=30) that corroborated and extended the behavioural findings. Two further MEG experiments planned for the near future, will exploit a similar paradigm but this time both perceptual features will be task relevant (i.e. a classical multiattribute task). One of the planned experiments will involve the modulation of various neurotransmitters via pharmacological manipulations in healthy participants.
Notably, due the limited data collection capacity during the pandemic, my team and I have re-analysed open-access behavioural data in previously published experiments that demonstrated “distractor” effects. Using detailed computational modelling and, more recently, by conducting our own behavioural experiments, we showed that some of the previously published distractor effects can be ascribed to simple reinforcement learning mechanisms. These findings are currently being prepared for publication.
The explorations and technical developments above were exploited in the first MEG experiment (N=20, 2 5-hour sessions per participant). In this experiment we presented to participants 3 Gabor patches against grey background, whose contrast levels and orientations varied, and across different trials asked them to select either the “highest” or the “lowest” contrast. Thus, unlike multiattribute decisions in which both dimensions are relevant, in this paradigm orientation was an irrelevant dimension. At the behavioural level we observed the so-called “distractor” effect, previously occurred in unidimensional decisions (according to which the higher the value of the worst alternative is the lower the probability of telling apart the two best alternatives is). At the same time, and despite the fact that orientation was in this experiment irrelevant, we observed well-known multiattribute contextual preference reversal effects (e.g. the attraction effect). These findings indicate that the perceptual (dis-) similarity among alternatives, even on the basis of irrelevant features, biases the way information is sampled and processed. At the neural level, we are currently focusing on how the decoded sensory representation associated with the three alternatives changes in the course of a trial a) as a function of the task framing (“highest” vs. “lowest”), b) as a function of the value of the worst alternative.
The above-mentioned MEG experiment was coupled with a similar behavioural experiment (N=30) that corroborated and extended the behavioural findings. Two further MEG experiments planned for the near future, will exploit a similar paradigm but this time both perceptual features will be task relevant (i.e. a classical multiattribute task). One of the planned experiments will involve the modulation of various neurotransmitters via pharmacological manipulations in healthy participants.
Notably, due the limited data collection capacity during the pandemic, my team and I have re-analysed open-access behavioural data in previously published experiments that demonstrated “distractor” effects. Using detailed computational modelling and, more recently, by conducting our own behavioural experiments, we showed that some of the previously published distractor effects can be ascribed to simple reinforcement learning mechanisms. These findings are currently being prepared for publication.
This work brings together decision-making and visual attention theories in an attempt to explain why humans behave irrationally. The work performed so far advanced our understanding of what drives so-called distractor effects. Previous explanations of this effect relied on a theory of neural coding dubbed “divisive normalisation”. Our findings thus far suggest that, even divisive normalisation moulds sensory representation early-on, attentional modulation and information sampling overrides such early representational distortions. The planned experiments will dig further into classical multiattribute contextual preference reversals, that have been challenging decision theories for decades. Additionally, we will link these behavioural phenomena to the function various neurotransmitters aiming to develop a neurophysiologically detailed theory of multiattribute choice–from the level of neurotransmitters, to large-scale brain networks, to behaviour.
At the methodological level our work has exploited and tested state-of-the-art techniques towards reading out the flow of information during challenging decision-making tasks involving three alternatives. This approach is expected to be inspiring and useful to researchers interested in probing complex behaviour at a fine temporal scale, using whole brain neuroimaging.
At the methodological level our work has exploited and tested state-of-the-art techniques towards reading out the flow of information during challenging decision-making tasks involving three alternatives. This approach is expected to be inspiring and useful to researchers interested in probing complex behaviour at a fine temporal scale, using whole brain neuroimaging.