Visual Search

Talk Session: Wednesday, May 24, 2023, 8:15 – 10:00 am, Talk Room 2
Moderator: Monica Castelhano, University British Columbia

Talk 1, 8:15 am, 61.21

Don’t hide the instruction manual: A dynamic trade-off between using internal and external templates during visual search

Alex Hoogerbrugge1, Christoph Strauch1, Tanja Nijboer1, Stefan Van der Stigchel1; 1Experimental Psychology, Helmholtz Institute, Utrecht University, The Netherlands

Visual search is typically studied by requiring participants to memorize a template initially, for which they subsequently search in a crowded display. Search in daily life, however, often involves templates that remain accessible externally, and may therefore be (re)attended for just-in-time encoding or to boost internal template representations. Here, we show that participants indeed use external templates during search when given the chance. This behavior was observed during both simple and complex search, scaled with task difficulty, and was associated with improved performance. We conclude that the external world may not only provide the challenge (e.g., distractors), but may dynamically ease search. These results argue for the extension of state-of-the-art models of search, as external sampling seems to be the default option and is actually beneficial for behavior. Our findings support a model of visual working memory that emphasizes a resource-efficient trade-off between storing and (re)accessing external information.

Acknowledgements: This work was supported by ERC [ERC-CoG-863732], https://erc.europa.eu/, awarded to Stefan Van der Stigchel.

Talk 2, 8:30 am, 61.22

Memory Compression Facilitates Search for Multiple Targets

Andrew Clement1 (), Brian Anderson1; 1Texas A&M University

Searching for multiple targets is often less efficient than searching for a single target. However, statistical learning and other higher-order regularities have been found to compress object representations in memory, enabling more efficient storage of information. Here, we tested whether this form of memory compression can facilitate search for multiple targets. Participants searched for one of two targets that were cued on each trial. In an initial training phase, specific pairs of targets always co-occurred with each other. In a subsequent test phase, participants searched for previously associated or unassociated pairs of targets. Participants were faster and more accurate at detecting the target when they searched for a previously associated pair of targets. Thus, statistical learning facilitated search for multiple targets. In a second experiment, we tested whether these effects were due to search guidance or decision-making processes. Participants completed the same task while we recorded their eye movements. Again, participants were faster and more accurate at detecting the target when they searched for a previously associated pair of targets. Participants were also more likely to initially fixate the target and fixated fewer distractors, suggesting that statistical learning facilitated search guidance. However, similar findings were not observed for dwell times on the target or distractors, suggesting that these findings were not due to target identification or distractor rejection processes. In a third experiment, we tested whether these effects generalized to other higher-order regularities. Participants completed a similar task in which they searched for semantically related or unrelated pairs of targets. Participants were faster and more accurate at detecting the target when they searched for a semantically related pair of targets. Thus, these effects generalized to semantic relationships. Together, these findings suggest that statistical learning and other higher-order regularities can facilitate search for multiple targets by compressing target representations in memory.

Talk 3, 8:45 am, 61.23

Looking for details: Fine-grained visual search at foveal scale

Sanjana Kapisthalam1 (), Martina Poletti1; 1University of Rochester, 2Center for vision science at University of Rochester

The concept of visual search is normally associated with large saccades, which are used to place the high-resolution fovea on the stimuli of interest. Yet, during fixation, fine spatial detail at the center of gaze is actively sampled through microsaccades. Using high-resolution eye-tracking coupled with a gaze-contingent display control system, enabling localization of the line of sight with arcminute precision, we investigated if the visuomotor system can engage in visual search at a finer scale during fixation. Subjects (n=9) were instructed to localize a target (a tilted bar: 1x8 arcminutes in size) in an array of 7 similar items. The stimulus array, presented for 1s, spanned 0.5 degrees, approximately half the size of the foveola. Salience was modulated by changing the color of the stimuli, whereas task relevance was modulated by changing their degree of tilt. Despite stimuli being presented foveally, our results show that subjects did not maintain fixation at the center of the stimulus array, but engaged in active visual search, using microsaccades as small as 10 arcminutes. Microsaccades benefited performance; in the absence of microsaccades the probability of correctly localizing the target dropped by 10% (p=0.01). Perceptual saliency influenced microsaccades, which were more precise and their latency was 120ms faster in the presence of a salient target. However, when the distractor was salient, microsaccades towards the distractor were actively suppressed. Microsaccade precision was also influenced by the task-relevance of distractors; when the distractors’ orientation was similar to that of the target the probability of microsaccades landing on the target dropped (0.55 vs 0.25; p=0.01). These findings suggest that the visuomotor system can engage in a fine-grained visual search and is capable of establishing high-resolution priority maps of the foveal space based on the relevance and salience of individual details.

Acknowledgements: Funded by Meta, Inc

Talk 4, 9:00 am, 61.24

Alpha oscillations in early visual cortex support visual search through inhibition of neuronal excitability to Target and Distractor features

Katharina Duecker1 (), Kimron L Shapiro1, Simon Hanslmary2, Jeremy Wolfe3,4, Yali Pan1, Ole Jensen1; 1Centre for Human Brain Health, School of Psychology, University of Birmingham, UK, 2Centre for Cognitive Neuroimaging, School of Neuroscience and Psychology, University of Glasgow, UK, 3Brigham and Women’s Hospital, Boston, MA USA, 4Harvard Medical School, Boston, Massachusetts, USA

Visual search models typically employ priority maps guiding attention towards Targets and away from Distractors. Ventral stream neurons have been shown to respond more strongly to Targets than Distractors, however it is debated whether neurons in early visual regions are similarly modulated. Neuronal alpha oscillations have long been suggested to inhibit Distractors, yet it has not been established if they contribute to the priority map. Here, we show that feature guidance modulates neuronal excitability in early visual regions (V1/V2). Importantly, we demonstrate that alpha oscillations in V1 facilitate search through functional inhibition. These results were obtained by exploiting the good spatio-temporal resolution of MEG and a novel high-frequency tagging approach (Rapid Invisible Frequency Tagging, RIFT). Using RIFT, we probed the neuronal excitability to a blue or yellow Target among blue and yellow Distractors in a classic visual search paradigm, when colour was a guiding feature or was irrelevant (unguided search). The colours were tagged at 60 and 67 Hz, respectively, making the flicker unperceivable. The RIFT responses (in V1/V2) to the Target colour were significantly enhanced whereas responses to Distractors were reduced in guided compared to unguided search. Importantly, strong alpha power in the early visual cortex before the onset of the search display predicted faster search times for guided and unguided searches. Alpha power also predicted reduced RIFT responses to all stimuli. Our results link feature guidance to gain modulation in V1/V2. This suggests that the priority map affects neuronal activity in early visual regions. Furthermore, we show that alpha oscillations in V1 are associated with enhanced performance and reduced neuronal excitability. We propose that functional inhibition by alpha oscillations sets a threshold on all stimuli. As alpha power decreases with the search display onset, it allows the boosted Targets to overcome the inhibition, enabling a more efficient search.

Acknowledgements: This research was supported by a Wellcome Trust Investigator Award in Science 207550

Talk 5, 9:15 am, 61.25

Rank and career level are inadequate measures of perceptual expertise in radiology

Robert G. Alexander1 (), Stephen Waite1, Shawn Lyo1, Ashwin Venkatakrishnan1, Arcadij Grigorian1, Stephen L. Macknik1, Susana Martinez-Conde1; 1SUNY Downstate Health Sciences University, Brooklyn, NY, USA

Radiology professionals and trainees are often classified into broad categories of expertise that are inferred from their title, level of training (i.e., intern, resident, attending, specialist), and/or years of experience—rather than by any objective metric of performance. These categories dramatically oversimplify reality, overlooking individual differences, exceptional skills or natural talent, and potential age-related declines in sensitivity. Rank is usually tied to years on the job, and individuals typically move up but not down the ladder, even if their skills diminish over time. A more principled measure of perceptual expertise would provide the basis to optimize training assessments by allowing medical education programs to directly test whether trainees are behaving like experts and—if not—what specific behaviors are not yet at a desired level of performance. We conducted a psychophysical and eye-tracking study aimed at quantifying the gaze dynamics used by professional radiologists to detect abnormalities in medical images. Naive individuals with no medical imaging experience (n=9), radiology residents (n=11), and attending radiologists (n=6) searched through chest X-rays, each of which contained one abnormality (a potentially cancerous nodule). Consistent with prior work, we found that some observers performed better than expected based on rank and years of experience. In fact, some residents early in their training outperformed attending radiologists, despite an extensive experience differential. These results highlight that, at best, experience is an uncertain predictor of expertise level, and at worse, it reflects little more than seniority. We therefore propose that individuals should instead be grouped based on their objectively measured performance in specific tasks. There is a need for the radiology field to move beyond the standard rank descriptions of attendings versus trainees, and to develop more efficient strategies and methods to quantify expertise.

Acknowledgements: This work was supported by the New York State Empire Innovator Program, by the National Science Foundation (Award 1734887 to SM-C and SLM; Award 1523614 to SLM), and by the National Institute of Health (Award R01EY031971 to SM-C and SLM; Award R01CA258021 to SM-C, SLM, and SW).

Talk 6, 9:30 am, 61.26

Expectations Versus Reality: The Effect of Semantic Knowledge on Statistical Learning

Laura Sikun Li1 (), Hannah Lum Smith1, Karolina Krzyƛ1, Carrick C. Williams2, Monica S. Castelhano1; 1Queen's University, 2California State University San Marcos

Throughout our daily lives, we are constantly perceiving a flow of visual information. Previous research has shown that we extract frequencies and similarities (i.e., statistical learning) to help us understand and organise incoming information. Statistical learning can be used to increase efficiency in visual search tasks (Jiang & Sisk, 2019), but it is unclear how that learning is affected by prior knowledge about the world. In the present study, we investigated the effect of semantic knowledge on statistical learning. The experiment consisted of a learning and memory phase. In the learning phase, 129 participants were presented with four objects on a blank background (one per quadrant) and were asked to search for a target object, cued by its picture. Targets appeared in either high or low probability locations (80% or 20% of trials, respectively). Of theoretical interest, targets were placed in either semantically consistent (e.g., basketball net in upper quadrants) or inconsistent locations (e.g., chandelier in lower quadrants). In the memory phase, participants indicated where each target most often appeared (i.e., the high probability location). We found that participants had slower response times for low compared to high probability locations, which supports previous research. Critically, we found a significant interaction, where participants responded faster overall to targets in semantically consistent than inconsistent locations. These results indicated that where these objects typically appear in the world influenced learning, even when displayed without context. In the memory phase, we found that although accuracy was high for both, it was significantly higher for consistent (87%) than inconsistent (84%) high probability locations. Overall, these findings suggest that learning does not occur in a vacuum. Despite the short learning window, semantic knowledge had a significant influence on learning and performance.

Talk 7, 9:45 am, 61.27

Stop pretending your trials are independent: Learn more from your data with asymptotic regression

Alasdair Clarke1 (), Amelia Hunt2; 1Department of Psychology, University of Essex, 2School of Psychology, University of Aberdeen

Typical experiments in vision science include a brief set of practise trials to ensure participants understand the task, and then a series of trials that repeat a set of conditions multiple times. The unit of analysis is typically a measure of central tendency taken from all the trials within a given condition/participant, which represents an estimate of performance in that condition that can be compared to other conditions. But inherent in any set of repeated trials is the way performance changes; for example, stimulus-response mappings get faster and less error-prone, and participants tune their attention to pick up information from the appropriate locations and start to anticipate sequences of events and their timing with more precision. We throw all this information away when we express performance as an average. Multi-level models (LMM) are an increasingly popular approach, but these assume that all trials are statistically independent and identically distributed. Here we advocate for applying a straightforward asymptotic regression (Stevens, 1951) to repeated-measures performance data and show how it provides a richer and more accurate measure of the effects of different manipulations on performance. Methods based on "average performance" (whether using aggregate statistics or a multi-level framework) tend to result in statistics that are biased by the early trials. We demonstrate the utility of asymptotic regression using data from classic paradigms such as visual search and show the extent to which established effects are driven by a) performance in early trials b) differences in the stable performance asymptote and c) how the learning rate varies between conditions. We think asymptotic regression should become a standard tool applied to repeated measures data to provide a richer picture of the dynamic changes in performance inherent in experiments with repeated measures.

Acknowledgements: This research was funded by the Economic and Social Research Council grant number ES/S016120/1 to A.D.F.C. and A.R.H.