S3 – How can you be so sure? Behavioral, computational, and neuro-scientific perspectives on metacogni-tion in perceptual decision-making
Time/Room: Friday, May 19, 2017, 2:30 – 4:30 pm, Talk Room 1
Organizer(s): Megan Peters, University of California Los Angeles
Presenters: Megan Peters, Ariel Zylberberg, Michele Basso, Wei Ji Ma, Pascal Mamassian
Metacognition, or our ability to monitor the uncertainty of our thoughts, decisions, and perceptions, is of critical importance across many domains . Here we focus on metacognition in perceptual decisions — the continuous inferences that we make about the most likely state of the world based on incoming sensory information. How does a police officer evaluate the fidelity of his perception that a perpetrator has drawn a weapon? How does a driver compute her certainty in whether a fleeting visual percept is a child or a soccer ball, impacting her decision to swerve? These kinds of questions are central to daily life, yet how such ‘confidence’ is computed in the brain remains unknown. In recent years, increasingly keen interest has been directed towards exploring such metacognitive mechanisms from computational (e.g., Rahnev et al., 2011, Nat Neuro; Peters & Lau, 2015, eLife), neuroimaging (e.g., Fleming et al., 2010, Science), brain stimulation (e.g., Fetsch et al., 2014, Neuron), and neuronal electro-physiology (e.g., Kiani & Shadlen, 2009, Science; Zylberberg et al., 2016, eLife) perspectives. Importantly, the computation of confidence is also of increasing interest to the broader range of researchers studying the computations underlying perceptual decision-making in general. Our central focus is on how confidence is computed in neuronal populations, with attention to (a) whether perceptual decisions and metacognitive judgments depend on the same or different computations, and (b) why confidence judgments sometimes fail to optimally track the accuracy of perceptual decisions. Key themes for this symposium will include neural correlates of confidence, behavioral consequences of evidence manipulation on confidence judgments, and computational characterizations of the relationship between perceptual decisions and our confidence in them. Our principal goal is to attract scientists studying or interested in confidence/uncertainty, sensory metacognition, and perceptual decision-making from both human and animal perspectives, spanning from the computational to the neurobiological level. We bring together speakers from across these disciplines, from animal electrophysiology and behavior through computational models of human uncertainty, to communicate their most recent and exciting findings. Given the recency of many of the findings discussed, our symposium will cover terrain largely untouched by the main program. We hope that the breadth of research programs represented in this symposium will encourage a diverse group of scientists to attend and actively participate in the discussion.
Transcranial magnetic stimulation to visual cortex induces suboptimal introspection
Speaker: Megan Peters, University of California Los Angeles
Additional Authors: Megan Peters, University of California Los Angeles; Jeremy Fesi, The Graduate Center of the City University of New York; Namema Amendi, The Graduate Center of the City University of New York; Jeffrey D. Knotts, University of California Los Angeles; Hakwan Lau, UCLA
In neurological cases of blindsight, patients with damage to primary visual cortex can discriminate objects but report no visual experience of them. This form of ‘unconscious perception’ provides a powerful opportunity to study perceptual awareness, but because the disorder is rare, many researchers have sought to induce the effect in neurologically intact observers. One promising approach is to apply transcranial magnetic stimulation (TMS) to visual cortex to induce blindsight (Boyer et al., 2005), but this method has been criticized for being susceptible to criterion bias confounds: perhaps TMS merely reduces internal visual signal strength, and observers are unwilling to report that they faintly saw a stimulus even if they can still discriminate it (Lloyd et al., 2013). Here we applied a rigorous response-bias free 2-interval forced-choice method for rating subjective experience in studies of unconscious perception (Peters and Lau, 2015) to address this concern. We used Bayesian ideal observer analysis to demonstrate that observers’ introspective judgments about stimulus visibility are suboptimal even when the task does not require that they maintain a response criterion — unlike in visual masking. Specifically, observers appear metacognitively blind to the noise introduced by TMS, in a way that is akin to neurological cases of blindsight. These findings are consistent with the hypothesis that metacognitive judgments require observers to develop an internal model of the statistical properties of their own signal processing architecture, and that introspective suboptimality arises when that internal model abruptly becomes invalid due to external manipulations.
The influence of evidence volatility on choice, reaction time and confidence in a perceptual decision
Speaker: Ariel Zylberberg, Columbia University
Additional Authors: Ariel Zylberberg, Columbia University; Christopher R. Fetsch, Columbia University; Michael N. Shadlen, Columbia University
Many decisions are thought to arise via the accumulation of noisy evidence to a threshold or bound. In perceptual decision-making, the bounded evidence accumulation framework explains the effect of stimulus strength, characterized by signal-to-noise ratio, on decision speed, accuracy and confidence. This framework also makes intriguing predictions about the behavioral influence of the noise itself. An increase in noise should lead to faster decisions, reduced accuracy and, paradoxically, higher confidence. To test these predictions, we introduce a novel sensory manipulation that mimics the addition of unbiased noise to motion-selective regions of visual cortex. We verified the effect of this manipulation with neuronal recordings from macaque areas MT/MST. For both humans and monkeys, increasing the noise induced faster decisions and greater confidence over a range of stimuli for which accuracy was minimally impaired. The magnitude of the effects was in agreement with predictions of a bounded evidence accumulation model.
A role for the superior colliculus in decision-making and confidence
Speaker: Michele Basso, University of California Los Angeles
Additional Authors: Michele Basso, University of California Los Angeles; Piercesare Grimaldi, University of California Los Angeles; Trinity Crapse, University of California Los Angeles
Evidence implicates the superior colliculus (SC) in attention and perceptual decision-making . In a simple target-selection task, we previously showed that discriminability between target and distractor neuronal activity in the SC correlated with decision accuracy, consistent with the hypothesis that SC encodes a decision variable. Here we extend these results to determine whether SC also correlates with decision criterion and confidence. Trained monkeys performed a simple perceptual decision task in two conditions to induce behavioral response bias (criterion shift): (1) the probability of two perceptual stimuli was equal, and (2) the probability of one perceptual stimulus was higher than the other. We observed consistent changes in behavioral response bias (shifts in decision criterion) that were directly cor-related with SC neuronal activity. Furthermore, electrical stimulation of SC mimicked the effect of stimulus probability manipulations, demonstrating that SC correlates with and is causally involved in setting decision criteria. To assess confidence, monkeys were offered a ‘safe bet’ option on 50% of trials in a similar task. The ‘safe bet’ always yielded a small reward, encouraging monkeys to select the ‘safe bet’ when they were less confident rather than risk no reward for a wrong decision. Both monkeys showed metacognitive sensitivity: they chose the ‘safe bet’ more on more difficult trials. Single- and multi-neuron recordings from SC revealed two distinct neuronal populations: one that discharged more robustly for more confident trials, and one that did so for less confident trials. Together these finding show how SC encodes information about decisions and decisional confidence.
Testing the Bayesian confidence hypothesis
Speaker: Wei Ji Ma, New York University
Additional Authors: Wei Ji Ma, New York University; Will Adler, New York University; Ronald van den Berg, University of Uppsala
Asking subjects to rate their confidence is one of the oldest procedures in psychophysics. Remarkably, quantitative models of confidence ratings have been scarce. What could be called the “Bayesian confidence hypothesis” states that an observer’s confidence rating distribution is completely determined by posterior probability. This hypothesis predicts specific quantitative relationships between performance and confidence. It also predicts that stimulus combinations that produce the same posterior will also produce the same confidence distribution. We tested these predictions in three contexts: a) perceptual categorization; b) visual working memory; c) the interpretation of scientific data.
Integration of visual confidence over time and across stimulus dimensions
Speaker: Pascal Mamassian, Ecole Normale Supérieure
Additional Authors: Pascal Mamassian, Ecole Normale Supérieure; Vincent de Gardelle, Université Paris 1; Alan Lee, Lingnan University
Visual confidence refers to our ability to estimate our own performance in a visual decision task. Several studies have highlighted the relatively high efficiency of this meta-perceptual ability, at least for simple visual discrimination tasks. Are observers equally good when visual confidence spans more than one stimulus dimension or more than a single decision? To address these issues, we used the method of confidence forced-choice judgments where participants are prompted to choose between two alter-natives the stimulus for which they expect their performance to be better (Barthelmé & Mamassian, 2009, PLoS CB). In one experiment, we asked observers to make confidence choice judgments between two different tasks (an orientation-discrimination task and a spatial-frequency-discrimi-nation task). We found that participants were equally good at making these across-dimensions confidence judgments as when choices were restricted to a single dimension, suggesting that visual confidence judgments share a common currency. In another experiment, we asked observers to make confidence-choice judgments between two ensembles of 2, 4, or 8 stimuli. We found that participants were increasingly good at making ensemble confidence judgments, suggesting that visual confidence judgments can accumulate information across several trials. Overall, these results help us better understand how visual confidence is computed and used over time and across stimulus dimensions.