Biology subjects

Rahnev, D.

Publications and source records attributed to Rahnev, D..

At least 19 recordsLinked to original sources

bioRxiv2026

Standardized human tests show that most vision models are face-blind

Face recognition ability varies enormously across humans. Individuals with face-blindness (prosopagnosia) struggle to recognize even close family members while super-recognizers can identify strangers with exceptional accuracy. Where do ANNs fall within the human face recognition spectrum? Here, we administered the standardized tests used to characterize human face recognition ability to a diverse set of models, allowing us to contextualize a model's performance within the distribution of human behavior. We found that the majority 55% of models to be classified as face-blind and that even the best face-trained models do not cross the human threshold to be considered a super-recognizer. Interrogating the internal representations of these models showed that models that performed well on standardized face recognition tasks were more identity-selective and viewpoint-invariant. We also found that low-performing models did contain some identity information in independent representational subspaces. Removing viewpoint-dependent subspace improved face recognition abilities in 49 of the 53 models tested. Targeted unit ablations further identified opposing contributions, with viewpoint-dependent units disrupting identity coding and viewpoint-tolerant units supporting it. Together, our results show that most AI models are face-blind with worse face recognition ability than humans, and demonstrate how differences across AI models can be used generate testable hypotheses about the computational basis of human face recognition in humans which can then be probed in future studies.

neuroscience↗

bioRxiv2026

Bias-aware versus bias-blind confidence in humans and machines

Confidence evaluates the likely accuracy of a current decision. However, to be maximally informative about accuracy, confidence judgments should incorporate information about ones broader decision tendencies, such as their propensity to favor specific alternatives. We distinguish bias-aware confidence, which considers such tendencies, from bias-blind confidence, which relies only on evidence available on the current trial. To adjudicate between bias-aware and bias-blind confidence, we identified a signature of bias-aware confidence: the down-weighting of confidence for alternatives that a participant is biased toward. We then used a large dataset (N = 200) spanning 4- and 8-choice digit-classification tasks to show that humans reliably exhibit this signature of bias-aware confidence. This effect was reduced under speed pressure and could not be explained by guessing. In contrast to the human results, artificial neural networks (ANNs) trained for object recognition lacked this signature of bias- aware confidence. Importantly, augmenting ANNs with a metacognitive module that allows confidence to take the networks biases into account led to the emergence of human-like bias- aware confidence. These findings show that human confidence incorporates not only information from the current trial but also longer-term decision tendencies, and that this capacity - absent in standard ANNs - can be conferred through specialized metacognitive mechanisms.

neuroscience↗

bioRxiv2026

Individual differences in artificial neural networks capture individual differences in human behavior

Human behavior differs substantially across individuals. While artificial neural networks (ANNs) are regarded as promising models of human perception, they are often assumed to lack such individual differences. Here, we demonstrate that multiple instances of the same ANN architecture exhibit substantial individual differences in behavior that mimic those observed in humans. We trained and tested 60 ANN instances from three architectures on a digit recognition task and found notable individual differences in overall accuracy, confidence, and response time (RT). Critically, these individual differences in ANN instances mapped consistently onto the individual differences produced by 60 humans performing the same task, with the mapping strength often approaching the human-to-human benchmark across all three behavioral metrics (accuracy, confidence, RT). The mapping generalized even across behavioral metrics: an ANN instance that aligned with an individual human on accuracy also aligned with the same individual on confidence and RT. These findings generalized to a more complex, 10-choice blurry object recognition task, though the human-ANN mapping was generally less robust than the human-human benchmark. Overall, these findings open the possibility of using ANN ensembles as computational proxies for probing the mechanisms underlying human variability.

neuroscience↗

bioRxiv2025

Human-like individual differences emerge from random weight initializations in neural networks

Much of AI research targets the behavior of an average human, a focus that traces to Turings imitation game. Yet, no two human individuals behave exactly alike. In this study, we show that artificial neural networks (ANNs) trained with different random initializations exhibit substantial individual differences that resemble those in humans. Using a large dataset (N = 60) of human responses (accuracy, confidence, & response time) in a digit recognition task, we trained multiple instances of three ANN architectures on the same task, creating as many ANN instances as human subjects. We found that these ANN instances vary significantly from one another. Critically, ANN instances showed consistent variation in their alignment with specific human subjects. This consistency in alignment between ANN instances and humans extended across behavioral metrics, indicating that an ANN instance mimicking an individual on one metric also does so on others. Finally, we showed that leveraging these alignments improves predictions of individual human responses. Our findings highlight the potential of ANNs to capture human variability, opening new directions to develop models that go beyond aligning the average human and instead aligning the idiosyncratic behavior of specific individuals.

neuroscience↗

bioRxiv2024

Is subjective perceptual similarity metacognitive?

Perceptual similarity is a cornerstone for human learning and generalization. However, in assessing perceptual similarity between two stimuli, it is not well-defined which features or combinations of features one should focus on. The problem has accordingly been considered ill-posed, implying a lack of a clear ground-truth basis for similarity judgments. We hypothesized that the stimuli rated as subjectively similar are those that are, in fact, more challenging for oneself to discern in practice, in near-threshold settings (e.g., psychophysics experiments). Therefore, an individuals perceptual discrimination capacities can serve as a quasi-objective ground truth for determining whether two stimuli should be judged as similar. To test this idea, we asked participants to rank the similarity of face pairs and also separately measured their discrimination capacity for those pairs. We found a positive association between perceptual discrimination capacity and subjective perceptual dissimilarity, with this association being specific to each individual. The results indicate that perceptual similarity judgments reflect and predict ones own perceptual discrimination capacities. Overall, although subjective, similarity judgments are normative in that they are grounded in ones perceptual capacities, which can be measured through a psychophysical task with objective responses.

neuroscience↗

bioRxiv2024

Behavior can be decoded across the cortex when individual differences are considered

Group-level analyses have typically linked behavioral signatures to a constrained set of brain areas. Here we show that two behavioral metrics - reaction time (RT) and confidence - can be decoded across the cortex when each individual is considered separately. Subjects (N=50) completed a perceptual decision-making task with confidence. We built models decoding trial-level RT and confidence separately for each subject using the activation patterns in one brain area at a time after splitting the entire cortex into 200 regions of interest (ROIs). First, we developed a simple test to determine the robustness of decoding performance, which showed that several hundred trials per subject are required for robust decoding. We then examined the decoding performance at the group and subject levels. At the group level, we replicated previous results by showing that both RT and confidence could be decoded from a small number of ROIs (12.0% and 3.5%, respectively). Critically, at the subject level, both RT and confidence could be decoded from most brain regions even after Bonferroni correction (90.0% and 72.5%, respectively). Surprisingly, we observed that many brain regions exhibited opposite brain-behavior relationships across individuals, such that, for example, higher activations predicted fast RTs in some subjects but slow RTs in others. All results were replicated in a second dataset. These findings show that behavioral signatures can be decoded from a much broader range of cortical areas than previously recognized and suggest the need to study the brain-behavior relationship at both the group and subject levels.

neuroscience↗

bioRxiv2024

Human-like dissociations between confidence and accuracy in convolutional neural networks

Prior research has shown that manipulating stimulus energy by changing both stimulus contrast and variability results in confidence-accuracy dissociations in humans. Specifically, even when performance is matched, higher stimulus energy leads to higher confidence. The most common explanation for this effect is the positive evidence heuristic where confidence neglects evidence that disconfirms the choice. However, an alternative explanation is the signal-and-variance-increase hypothesis, according to which these dissociations arise from low-level changes in the separation and variance of perceptual representations. Because artificial neural networks lack built-in confidence heuristics, they can serve as a test for the necessity of confidence heuristics in explaining confidence-accuracy dissociations. Therefore, we tested whether confidence-accuracy dissociations induced by stimulus energy manipulations emerge naturally in convolutional neural networks (CNNs). We found that, across three different energy manipulations, CNNs produced confidence-accuracy dissociations similar to those found in humans. This effect was present for a range of CNN architectures from shallow 4-layer networks to very deep ones, such as VGG-19 and ResNet -50 pretrained on ImageNet. Further, we traced back the reason for the confidence-accuracy dissociations in all CNNs to the same signal-and-variance increase that has been proposed for humans: higher stimulus energy increased the separation and variance of the CNNs internal representations leading to higher confidence even for matched accuracy. These findings cast doubt on the necessity of the positive evidence heuristic to explain human confidence and establish CNNs as promising models for adjudicating between low-level, stimulus-driven and high-level, cognitive explanations of human behavior.

animal behavior and cognition↗

bioRxiv2023

Threat expectation does not improve perceptual discrimination despite causing heightened priority processing in the frontoparietal network

Threat cues have been widely shown to elicit increased sensory and attentional neural processing. However, whether this enhanced recruitment leads to measurable behavioral improvements in perception is still in question. Here we adjudicate between two opposing theories: that threat cues do or do not enhance perceptual sensitivity. We created threat stimuli by pairing one direction of motion in a random dot kinematogram with an aversive sound. While in the MRI scanner, 46 subjects (both men and women) completed a cued (threat/safe/neutral) perceptual decision-making task where they indicated the perceived motion direction of each moving dots stimulus. We found strong evidence that threat cues did not increase perceptual sensitivity compared to safe and neutral cues. This lack of improvement in perceptual decision-making ability occurred despite the threat cue resulting in widespread increases in frontoparietal BOLD activity, as well as increased connectivity between the right insula and the frontoparietal network. These results call into question the intuitive claim that expectation automatically enhances our perception of threat, and highlight the role of the frontoparietal network in prioritizing the processing of threat-related environmental cues. Significance StatementThreatening information receives enhanced priority processing in the brain. Evidence of increased neural activity to threat has fostered the current view that such selective processing leads to a boost in perception, suggesting that motivationally relevant top-down effects can directly change what we see. In the real world, danger is often preceded by an environmental cue that predicts its imminent approach. Here we used an aversive conditioning paradigm to test whether threat cues can change subjects ability to visually distinguish between threat and safe stimuli. Our results provide strong evidence for the lack of an effect of threat expectation on perceptual sensitivity, supporting the theory that perception is impenetrable by top-down cognitive influences despite robust neural attentional priority.

neuroscience↗

bioRxiv2023

Common computations in automatic cue combination and metacognitive confidence reports

It is well known that sensory information from one modality can automatically affect judgments from a different sensory modality. However, it remains unclear what determines the strength of the influence of an irrelevant sensory cue from one modality on a perceptual judgment for a different modality. Here we test whether the strength of multisensory impact by an irrelevant sensory cue depends on participants objective accuracy or subjective confidence for that cue. We created visual motion stimuli with low vs. high overall motion energy, where high-energy stimuli yielded higher confidence but lower accuracy in a visual-only task. We then tested the impact of the low- and high-energy visual stimuli on auditory motion perception. We found that the high-energy visual stimuli influenced the auditory motion judgments more strongly than the low-energy visual stimuli, consistent with their higher confidence but contrary to their lower accuracy. A computational model assuming common principles underlying confidence reports and multisensory integration captured these effects. Our findings show that automatic multisensory integration follows subjective confidence rather than objective performance and suggest the existence of common computations across vastly different stages of perceptual decision making.

neuroscience↗

bioRxiv2023

Multiple brain activation patterns for the same task

Meaningful variation in internal states that impacts cognition and behavior remains challenging to discover and characterize. Here we leveraged trial-to-trial fluctuations in the brain-wide signal recorded using functional MRI to test if distinct sets of brain regions are activated on different trials when accomplishing the same task. Across three different perceptual decision-making experiments, we estimated the brain activations for each trial. We then clustered the trials based on their similarity using modularity-maximization, a data-driven classification method. In each experiment, we found multiple distinct but stable subtypes of trials, suggesting that the same task can be accomplished in the presence of widely varying brain activation patterns. Surprisingly, in all experiments, one of the subtypes exhibited strong activation in the default mode network, which is typically thought to decrease in activity during tasks that require externally focused attention. The remaining subtypes were characterized by activations in different task-positive areas. The default mode network subtype was characterized by behavioral signatures that were similar to the other subtypes exhibiting activation with task-positive regions. These findings demonstrate that the same perceptual decision-making task is accomplished through multiple brain activation patterns.

neuroscience↗

bioRxiv2023

The timing of confidence computations in human prefrontal cortex

Knowing when confidence computations take place is critical for building mechanistic understanding of the neural and computational bases of metacognition. Yet, even though substantial amount of research has focused on revealing the neural correlates and computations underlying human confidence judgments, very little is known about the timing of confidence computations. Subjects judged the orientation of a briefly presented visual stimulus and provided a confidence rating regarding the accuracy of their decision. We delivered single pulses of transcranial magnetic stimulation (TMS) at different times after stimulus presentation. TMS was delivered to either dorsolateral prefrontal cortex (DLPFC) in the experimental group or to vertex in the control group. We found that TMS to DLPFC, but not to vertex, led to increased confidence in the absence of changes to accuracy or metacognitive ability. Critically, equivalent levels of confidence increase occurred for TMS delivered between 200 and 500 ms after stimulus presentation. These results suggest that confidence computations occur during a broad window that begins before the perceptual decision has been fully made and thus provide important constraints for theories of confidence generation.

neuroscience↗

bioRxiv2023

Brain signatures indexing variation in internal processing during perceptual decision-making

Brain activity is highly variable even while performing the same cognitive task with consequences for performance. Discovering, characterizing, and linking variability in brain activity to internal processes has primarily relied on experimentally inducing changes (e.g., via attention manipulation) to identify neuronal and behavioral consequences or studying spontaneous changes in ongoing brain dynamics. However, changes in internal processing could arise from many factors, such as variation in strategy or arousal, that are independent of experimental conditions. Here we utilize a data-driven clustering method based on modularity-maximation to identify consistent spatial-temporal EEG activity patterns across individual trials and relate this activity to behavioral performance. Subjects (N = 25) performed a motion direction discrimination task with six interleaved levels of motion coherence. Modularity-maximization based clustering identified two discrete spatial-temporal clusters, or subtypes, of trials with different patterns of brain activity. Surprisingly, even though Subtype 1 occurred more frequently with lower motion coherence, it was nonetheless associated with faster response times. Computational modeling suggests that Subtype 1 was characterized by a lower threshold for reaching a decision. These results highlight trial-to-trial variability in decision processes usually masked to experimenters and provide a method for identifying endogenous brain state variability relevant to cognition and behavior. Highlights Brain activity is highly variable. We find multiple and distinct stimulus-driven patterns in EEG. With changes in decision-making and drift-diffusion parameters. These results suggest a new way to identify brain states relevant to behavior. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=178 HEIGHT=200 SRC="FIGDIR/small/523502v3_ufig1.gif" ALT="Figure 1"> View larger version (42K): org.highwire.dtl.DTLVardef@e6e448org.highwire.dtl.DTLVardef@95512eorg.highwire.dtl.DTLVardef@104a976org.highwire.dtl.DTLVardef@11191e2_HPS_FORMAT_FIGEXP M_FIG C_FIG

neuroscience↗

bioRxiv2022

Task learning is subserved by a domain-general brain network

One of the most important human faculties is the ability to acquire not just new memories but the capacity to perform entirely new tasks. However, little is known about the brain mechanisms underlying the learning of novel tasks. Specifically, it is unclear to what extent learning of different tasks depends on domain-general and/or domain-specific brain mechanisms. Here human subjects (N=45) learned to perform six new tasks while undergoing functional MRI. The different tasks required the engagement of perceptual, motor, and various cognitive processes (attention, expectation, speed-accuracy tradeoff, and metacognition). We found that a bilateral frontoparietal network was more active during the initial compared to the later stages of task learning, and that this effect was stronger for task variants requiring more new learning. Critically, the same frontoparietal network was engaged by all six tasks, demonstrating its domain generality. Finally, although task learning decreased the overall activity in the frontoparietal network, it increased the connectivity strength between the different nodes of that network. These results demonstrate the existence of a domain-general brain network whose activity and connectivity reflect learning for a wide variety of new tasks, and thus may underlie the human capacity for acquiring new abilities.

neuroscience↗

bioRxiv2022

RTNet: A neural network that exhibits the signatures of human perceptual decision making

Convolutional neural networks show promise as models of biological vision. However, their decision behavior, including the facts that they are deterministic and use equal number of computations for easy and difficult stimuli, differs markedly from human decision-making, thus limiting their applicability as models of human perceptual behavior. Here we develop a new neural network, RTNet, that generates stochastic decisions and human-like response time (RT) distributions. We further performed comprehensive tests that showed RTNet reproduces all foundational features of human accuracy, RT, and confidence and does so better than all current alternatives. To test RTNets ability to predict human behavior on novel images, we collected accuracy, RT, and confidence data from 60 human subjects performing a digit discrimination task. We found that the accuracy, RT, and confidence produced by RTNet for individual novel images correlated with the same quantities produced by human subjects. Critically, human subjects who were more similar to the average human performance were also found to be closer to RTNets predictions, suggesting that RTNet successfully captured average human behavior. Overall, RTNet is a promising model of human response times that exhibits the critical signatures of perceptual decision making.

neuroscience↗

bioRxiv2022

Idiosyncratic relation between human brain activity and behavior

Research in neuroscience often assumes universal neural mechanisms, but increasing evidence points towards sizeable individual differences in brain activations. What remains unclear is the extent of the idiosyncrasy and whether different types of analyses are associated with different levels of idiosyncrasy. Here we develop a new method for addressing these questions. The method consists of computing the within-subject reliability and subject-to-group similarity of brain activations and submitting these values to a computational model that quantifies the relative strength of group- and subject-level factors. We apply this method to a perceptual decision-making task (N=50) and find that activations related to task, reaction time (RT), and confidence are influenced equally strongly by group- and subject-level factors. Both group- and subject-level factors are dwarfed by a noise factor, though higher levels of smoothing increases their contributions relative to noise. Overall, our method allows for the quantification of group- and subject-level factors of brain activations and thus provides a more detailed understanding of the idiosyncrasy levels in brain activations.

neuroscience↗

bioRxiv2020

TMS alters multivoxel patterns in the absence of overall activity changes

Transcranial magnetic stimulation (TMS) has become one of the major tools for establishing the causal role of specific brain regions in perceptual, motor, and cognitive processes. Nevertheless, a persistent limitation of the technique is the lack of clarity regarding its precise effects on neural activity. Here, we examined the effects of TMS intensity and frequency on concurrently recorded blood-oxygen level-dependent (BOLD) signals at the site of stimulation. In two experiments, we delivered TMS to the dorsolateral prefrontal cortex in human subjects of both sexes. In Experiment 1, we delivered a series of pulses at high (100% of motor threshold) or low (50% of motor threshold) intensity, whereas in Experiment 2, we always used high intensity but delivered stimulation at four different frequencies (5, 8.33, 12.5, and 25 Hz). We found that the TMS intensity and frequency could be reliably decoded using multivariate analysis techniques even though TMS had no effect on overall BOLD activity at the site of stimulation in either experiment. These results provide important insight into the mechanisms through which TMS influences neural activity. SignificanceTranscranial magnetic stimulation (TMS) is a promising tool for the treatment of a number of neuropsychiatric disorders. However, its effectiveness is still impeded by an incomplete understanding of its neural effects. One fundamental unresolved issue is whether TMS leads to local changes in overall neural activity in the absence of a task. Here we performed two experiments where TMS was delivered inside an MRI scanner while brain activity was continuously monitored. We found converging evidence for the notion that TMS affects the pattern of local activity changes but does not lead to an overall increase in activity. These results help clarify the mechanisms of how TMS affects local neural activity.

neuroscience↗

bioRxiv2019

Awake reactivation and suppression after brief task exposure depend on task novelty

Previously learned information is known to be reactivated during periods of quiet wakefulness and such awake reactivation is considered to be a key mechanism for memory consolidation. We recently demonstrated that feature-specific awake reactivation occurs in early visual cortex immediately after extensive visual training on a novel task. To understand the exact role of awake reactivation, here we investigated whether such reactivation depends specifically on the task novelty. Subjects completed a brief visual task that was either novel or extensively trained on previous days. Replicating our previous results, we found that awake reactivation occurs for the novel task even after a brief learning period. Surprisingly, however, brief exposure to the extensively trained task led to \"awake suppression\" such that neural activity immediately after the exposure diverged from the pattern for the trained task. Further, subjects who had greater performance improvement showed stronger awake suppression. These results suggest that the brain operates different post-task processing depending on prior visual training.

neuroscience↗

bioRxiv2019

The nature of the perceptual representation for decision making

It is becoming widely appreciated that human perceptual decision making is suboptimal but the nature and origins of this suboptimality remain poorly understood. Most past research has employed tasks with two stimulus categories, but such designs cannot fully capture the limitations inherent in naturalistic perceptual decisions where choices are rarely between only two alternatives. We conducted four experiments with tasks involving multiple alternatives and used computational modeling to determine the decision-level representation on which the perceptual decisions were based. The results from all four experiments pointed to the existence of robust suboptimality such that most of the information in the sensory representation was lost during the transformation to a decision-level representation. These results reveal severe limits in the quality of decision-level representations for multiple alternatives and have strong implications about perceptual decision making in naturalistic settings.

neuroscience↗