Salience Models: A Computational Cognitive Neuroscience Review

Perception of motion salience shapes the emergence of collective motions

Despite the profound implications of self-organization in animal groups for collective behaviors, understanding the fundamental principles and applying them to swarm robotics remains incomplete. Here we propose a heuristic measure of perception of motion salience (MS) to quantify relative motion changes of neighbors from first-person view. Leveraging three large bird-flocking datasets, we explore how this perception of MS relates to the structure of leader-follower (LF) relations, and further perform an individual-level correlation analysis between past perception of MS and future change rate of velocity consensus. We observe prevalence of the positive correlations in real flocks, which demonstrates that individuals will accelerate the convergence of velocity with neighbors who have higher MS. This empirical finding motivates us to introduce the concept of adaptive MS-based (AMS) interaction in swarm model. Finally, we implement AMS in a swarm of ~102 miniature robots. Swarm experiments show the significant advantage of AMS in enhancing self-organization of the swarm for smooth evacuations from confined environments.

The Gaze of Schizophrenia Patients Captured by Bottom-up Saliency

Schizophrenia (SCHZ) notably impacts various human perceptual modalities, including vision. Prior research has identified marked abnormalities in perceptual organization in SCHZ, predominantly attributed to deficits in bottom-up processing. Our study introduces a novel paradigm to differentiate the roles of top-down and bottom-up processes in visual perception in SCHZ. We analysed eye-tracking fixation ground truth maps from 28 SCHZ patients and 25 healthy controls (HC), comparing these with two mathematical models of visual saliency: one bottom-up, based on the physical attributes of images, and the other top-down, incorporating machine learning. While the bottom-up (GBVS) model revealed no significant overall differences between groups (beta = 0.01, p = 0.281, with a marginal increase in SCHZ patients), it did show enhanced performance by SCHZ patients with highly salient images. Conversely, the top-down (EML-Net) model indicated no general group difference (beta = -0.03, p = 0.206, lower in SCHZ patients) but highlighted significantly reduced performance in SCHZ patients for images depicting social interactions (beta = -0.06, p < 0.001). Over time, the disparity between the groups diminished for both models. The previously reported bottom-up bias in SCHZ patients was apparent only during the initial stages of visual exploration and corresponded with progressively shorter fixation durations in this group. Our research proposes an innovative approach to understanding early visual information processing in SCHZ patients, shedding light on the interplay between bottom-up perception and top-down cognition.

An improved saliency model of visual attention dependent on image content

Many visual attention models have been presented to obtain the saliency of a scene, i.e., the visually significant parts of a scene. However, some mechanisms are still not taken into account in these models, and the models do not fit the human data accurately. These mechanisms include which visual features are informative enough to be incorporated into the model, how the conspicuity of different features and scales of an image may integrate to obtain the saliency map of the image, and how the structure of an image affects the strategy of our attention system. We integrate such mechanisms in the presented model more efficiently compared to previous models. First, besides low-level features commonly employed in state-of-the-art models, we also apply medium-level features as the combination of orientations and colors based on the visual system behavior. Second, we use a variable number of center-surround difference maps instead of the fixed number used in the other models, suggesting that human visual attention operates differently for diverse images with different structures. Third, we integrate the information of different scales and different features based on their weighted sum, defining the weights according to each component's contribution, and presenting both the local and global saliency of the image. To test the model's performance in fitting human data, we compared it to other models using the CAT2000 dataset and the Area Under Curve (AUC) metric. Our results show that the model has high performance compared to the other models (AUC = 0.79 and sAUC = 0.58) and suggest that the proposed mechanisms can be applied to the existing models to improve them.

The Oddity Detection in Diverse Scenes (ODDS) database: Validated real-world scenes for studying anomaly detection

Many applied screening tasks (e.g., medical image or baggage screening) involve challenging searches for which standard laboratory search is rarely equivalent. For example, whereas laboratory search frequently requires observers to look for precisely defined targets among isolated, non-overlapping images randomly arrayed on clean backgrounds, medical images present unspecified targets in noisy, yet spatially regular scenes. Those unspecified targets are typically oddities, elements that do not belong. To develop a closer laboratory analogue to this, we created a database of scenes containing subtle, ill-specified "oddity" targets. These scenes have similar perceptual densities and spatial regularities to those found in expert search tasks, and each includes 16 variants of the unedited scene wherein an oddity (a subtle deformation of the scene) is hidden. In Experiment 1, eight volunteers searched thousands of scene variants for an oddity. Regardless of their search accuracy, they were then shown the highlighted anomaly and rated its subtlety. Subtlety ratings reliably predicted search performance (accuracy and response times) and did so better than image statistics. In Experiment 2, we conducted a conceptual replication in which a larger group of naïve searchers scanned subsets of the scene variants. Prior subtlety ratings reliably predicted search outcomes. Whereas medical image targets are difficult for naïve searchers to detect, our database contains thousands of interior and exterior scenes that vary in difficulty, but are nevertheless searchable by novices. In this way, the stimuli will be useful for studying visual search as it typically occurs in expert domains: Ill-specified search for anomalies in noisy displays.

Between the Scenes

We constantly move our eyes to new information while inspecting a scene, but these patterns of eye movements change based on the task and goals of the observer. Inhibition of return (IOR) may facilitate visual search by reducing the likelihood of revisiting previously attended locations. However, IOR may present in any visual task, or it may be search-specific. We investigated the presence of IOR in foraging, memorization, change detection, and two versions of visual search. One version of search used a static search array that remained stable throughout the trial, but the second used a scene flickering paradigm similar to the change detection task. IOR was observed in both versions of visual search, memorization, and foraging, but not in change detection. Visual search and change detection both had temporal nonscene components, and we observed that IOR could be maintained despite the scene removal but only for search. Although IOR is maintained in scene coordinates, short disruptions to this scene are insufficient to completely remove the inhibitory tags. Finally, we compare return saccades in trials without a probe and observe fewer return saccades in tasks for which IOR was observed, providing further evidence that IOR might serve as a novelty drive.

The Spatial Leaky Competing Accumulator Model

The Leaky Competing Accumulator model (LCA) of Usher and McClelland is able to simulate the time course of perceptual decision making between an arbitrary number of stimuli. Reaction times, such as saccadic latencies, produce a typical distribution that is skewed toward longer latencies and accumulator models have shown excellent fit to these distributions. We propose a new implementation called the Spatial Leaky Competing Accumulator (SLCA), which can be used to predict the timing of subsequent fixation durations during a visual task. SLCA uses a pre-existing saliency map as input and represents accumulation neurons as a two-dimensional grid to generate predictions in visual space. The SLCA builds on several biologically motivated parameters: leakage, recurrent self-excitation, randomness and non-linearity, and we also test two implementations of lateral inhibition. A global lateral inhibition, as implemented in the original model of Usher and McClelland, is applied to all competing neurons, while a local implementation allows only inhibition of immediate neighbors. We trained and compared versions of the SLCA with both global and local lateral inhibition with use of a genetic algorithm, and compared their performance in simulating human fixation latency distribution in a foraging task. Although both implementations were able to produce a positively skewed latency distribution, only the local SLCA was able to match the human data distribution from the foraging task. Our model is discussed for its potential in models of salience and priority, and its benefits as compared to other models like the Leaky integrate and fire network.

Semantic object-scene inconsistencies affect eye movements, but not in the way predicted by contextualized meaning maps

Semantic information is important in eye movement control. An important semantic influence on gaze guidance relates to object-scene relationships: objects that are semantically inconsistent with the scene attract more fixations than consistent objects. One interpretation of this effect is that fixations are driven toward inconsistent objects because they are semantically more informative. We tested this explanation using contextualized meaning maps, a method that is based on crowd-sourced ratings to quantify the spatial distribution of context-sensitive “meaning” in images. In Experiment 1, we compared gaze data and contextualized meaning maps for images, in which objects-scene consistency was manipulated. Observers fixated more on inconsistent versus consistent objects. However, contextualized meaning maps did not assign higher meaning to image regions that contained semantic inconsistencies. In Experiment 2, a large number of raters evaluated image-regions, which were deliberately selected for their content and expected meaningfulness. The results suggest that the same scene locations were experienced as slightly less meaningful when they contained inconsistent compared to consistent objects. In summary, we demonstrated that — in the context of our rating task — semantically inconsistent objects are experienced as less meaningful than their consistent counterparts and that contextualized meaning maps do not capture prototypical influences of image meaning on gaze guidance.

Glimpse: A Gaze-Based Measure of Temporal Salience

Temporal salience considers how visual attention varies over time. Although visual salience has been widely studied from a spatial perspective, its temporal dimension has been mostly ignored, despite arguably being of utmost importance to understand the temporal evolution of attention on dynamic contents. To address this gap, we proposed Glimpse, a novel measure to compute temporal salience based on the observer-spatio-temporal consistency of raw gaze data. The measure is conceptually simple, training free, and provides a semantically meaningful quantification of visual attention over time. As an extension, we explored scoring algorithms to estimate temporal salience from spatial salience maps predicted with existing computational models. However, these approaches generally fall short when compared with our proposed gaze-based measure. Glimpse could serve as the basis for several downstream tasks such as segmentation or summarization of videos. Glimpse's software and data are publicly available.

Inhibition of return: An information processing theory of its natures and significance

Inhibition of return (IOR) is an inhibitory aftereffect of visuospatial orienting, typically resulting in slower responses to targets presented in an area that has been recently attended. Since its discovery, myriad research has sought to explain the causes and effects underlying this phenomenon. Here, we briefly summarize the history of the phenomenon, and describe the early work supporting the functional significance of IOR as a foraging facilitator. We then shine a light on the discordance in the literature with respect to mechanism-in particular the lack of theoretical constructs that can consistently explain innumerable dissociations. We then describe three diagnostics (central arrow targets, locus of slack logic and the psychological refractory period, and performance in speed-accuracy space) used to support our theory that there are two forms of inhibition of return-the form which is manifest being contingent upon the activation state of the reflexive oculomotor system. The input form, which operates to decrease the salience of inputs, is generated when the reflexive oculomotor system is suppressed; the output form, which operates to bias responding, is generated when the reflexive oculomotor system is not suppressed. Then, we subject a published data set, wherein inhibitory effects had been generated while the reflexive oculomotor system was either active or suppressed, to diffusion modelling. As we hypothesized, based on the aforementioned theory, the effects of the two forms of IOR were best accounted for by different drift diffusion parameters. The paper ends with a variety of suggestions for further research.

No Advantage for Separating Overt and Covert Attention in Visual Search

We move our eyes roughly three times every second while searching complex scenes, but covert attention helps to guide where we allocate those overt fixations. Covert attention may be allocated reflexively or voluntarily, and speeds the rate of information processing at the attended location. Reducing access to covert attention hinders performance, but it is not known to what degree the locus of covert attention is tied to the current gaze position. We compared visual search performance in a traditional gaze-contingent display, with a second task where a similarly sized contingent window is controlled with a mouse, allowing a covert aperture to be controlled independently by overt gaze. Larger apertures improved performance for both the mouse- and gaze-contingent trials, suggesting that covert attention was beneficial regardless of control type. We also found evidence that participants used the mouse-controlled aperture somewhat independently of gaze position, suggesting that participants attempted to untether their covert and overt attention when possible. This untethering manipulation, however, resulted in an overall cost to search performance, a result at odds with previous results in a change blindness paradigm. Untethering covert and overt attention may therefore have costs or benefits depending on the task demands in each case.

Temporal Limitations of the Standard Leaky Integrate and Fire Model

Itti and Koch's Saliency Model has been used extensively to simulate fixation selection in a variety of tasks from visual search to simple reaction times. Although the Saliency Model has been tested for its spatial prediction of fixations in visual salience, it has not been well tested for their temporal accuracy. Visual tasks, like search, invariably result in a positively skewed distribution of saccadic reaction times over large numbers of samples, yet we show that the leaky integrate and fire (LIF) neuronal model included in the classic implementation of the model tends to produce a distribution shifted to shorter fixations (in comparison with human data). Further, while parameter optimization using a genetic algorithm and Nelder-Mead method does improve the fit of the resulting distribution, it is still unable to match temporal distributions of human responses in a visual task. Analysis of times for individual images reveal that the LIF algorithm produces initial fixation durations that are fixed instead of a sample from a distribution (as in the human case). Only by aggregating responses over many input images do they result in a distribution, although the form of this distribution still depends on the input images used to create it and not on internal model variability.

What Neuroscientific Studies Tell Us about Inhibition of Return

An inhibitory aftermath of orienting, inhibition of return (IOR), has intrigued scholars since its discovery about 40 years ago. Since then, the phenomenon has been subjected to a wide range of neuroscientific methods and the results of these are reviewed in this paper. These include direct manipulations of brain structures (which occur naturally in brain damage and disease or experimentally as in TMS and lesion studies) and measurements of brain activity (in humans using EEG and fMRI and in animals using single unit recording). A variety of less direct methods (e.g., computational modeling, developmental studies, etc.) have also been used. The findings from this wide range of methods support the critical role of subcortical and cortical oculomotor pathways in the generation and nature of IOR.