Unsupervised foveal vision neural architecture with top-down attention.

Deep learning architectures are an extremely powerful tool for recognizing and classifying images. However, they require supervised learning and normally work on vectors of the size of image pixels and produce the best results when trained on millions of object images. To help mitigate these issues, we propose an end-to-end architecture that fuses bottom-up saliency and top-down attention with an object recognition module to focus on relevant data and learn important features that can later be fine-tuned for a specific task, employing only unsupervised learning. In addition, by utilizing a virtual fovea that focuses on relevant portions of the data, the training speed can be greatly improved. We test the performance of the proposed Gamma saliency technique on the Toronto and CAT 2000 databases, and the foveated vision in the large Street View House Numbers (SVHN) database. The results with foveated vision show that Gamma saliency performs at the same level as the best alternative algorithms while being computationally faster. The results in SVHN show that our unsupervised cognitive architecture is comparable to fully supervised methods and that saliency also improves CNN performance if desired. Finally, we develop and test a top-down attention mechanism based on the Gamma saliency applied to the top layer of CNNs to facilitate scene understanding in multi-object cluttered images. We show that the extra information from top-down saliency is capable of speeding up the extraction of digits in the cluttered multidigit MNIST data set, corroborating the important role of top down attention.

Pre-target alpha power predicts the speed of cued target discrimination.

The human visual system selects information from dense and complex streams of spatiotemporal input. This selection process is aided by prior knowledge of the features, location, and temporal proximity of an upcoming stimulus. In the laboratory, this knowledge is often conveyed by cues, preceding a task-relevant target stimulus. Response speed in cued selection tasks varies within and across participants and is often thought to index efficient selection of a cued feature, location, or moment in time. The present study used a reverse correlation approach to identify neural predictors of efficient target discrimination: Participants identified the orientation of a sinusoidal grating, which was presented in one hemifield following the presentation of bilateral visual cues that carried temporal but not spatial information about the target. Across different analytic approaches, faster target responses were predicted by larger alpha power preceding the target. These results suggest that heightened pre-target alpha power during a cue period may index a state that is beneficial for subsequent target processing. Our findings are broadly consistent with models that emphasize capacity sharing across time, as well as models that link alpha oscillations to temporal predictions regarding upcoming events.

Attention to a threat-related feature does not interfere with concurrent attentive feature selection.

Visual features associated with a task and those that predict noxious events both prompt selectively heightened visuocortical responses. Conflicting views exist regarding how the competition between a task-related and a threat-related feature is resolved when they co-occur in time and space. Utilizing aversive classical Pavlovian conditioning, we investigated the visuocortical representation of two simultaneously presented, fully overlapping visual stimuli. Isoluminant red and green random dot kinematogram (RDK) stimuli were flickered at distinct tagging frequencies (8.57 Hz, 12 Hz) to elicit distinguishable steady-state visual evoked potentials (ssVEPs). Occasional coherent motion events prompted a motor response (task) or predicted a noxious noise (threat). These events occurred either in the green (task cue), the red (threat cue), or in both RDKs simultaneously. In the initial habituation phase, participants responded to coherent motion of the green RDK with a key press, but no loud noise was presented at any time. Here, selective amplification was seen for the task-relevant (green) RDK, and interference was observed when both RDKs simultaneously showed coherent motion. Upon pairing the threat cue with the noxious noise in the subsequent acquisition phase, the threat cue-evoked ssVEP (red RDK) was also amplified, but this amplification did not interact with amplification of the task cue or alter the behavioral or visuocortical interference effect observed during simultaneous coherent motion. Although competing feature conjunctions resulted in interference in the visual cortex, the acquisition of a bias toward an individual threat-related feature did not result in additional cost effects.

Assessing the relationship between pupil diameter and visuocortical activity.

Visuocortical activity and pupil diameter both increase in tasks involving memory, attention, and physiological arousal. Thus, the question arises whether pupil dilation prompts a subsequent increase in visuocortical activity. In this study, we investigated the extent to which changes in visuocortical activity relate to changes in pupil diameter. The amplitude of the sustained visuocortical response to a flickering stimulus (i.e., steady-state visually evoked potential [ssVEP] power) was examined in 39 participants while pupil diameter was measured. To generalize across stimulus conditions, Gabor stimuli varied in brightness and ssVEP driving frequency. As expected, brighter stimuli prompted pupil constriction and larger ssVEP power. To determine whether momentary fluctuations in pupil size contribute to the ssVEP amplitude under conditions of constant luminance and frequency, the single-trial means from each measure were correlated and the shape of the pupil-diameter waveform related to the ssVEP amplitude time course, both within and between participants. Under constant conditions, changes in pupil diameter were not related to changes in ssVEP amplitude, at any luminance level or driving frequency. Findings suggest that pupil dilation does not systematically prompt subsequent changes in visuocortical activity, and thus is not a sufficient cause of visuocortical modulation in cognitive or affective tasks.

What does the dot-probe task measure? A reverse correlation analysis of electrocortical activity.

The dot-probe task is considered a gold standard for assessing the intrinsic attentive selection of one of two lateralized visual cues, measured by the response time to a subsequent, lateralized response probe. However, this task has recently been associated with poor reliability and conflicting results. To resolve these discrepancies, we tested the underlying assumption of the dot-probe task-that fast probe responses index heightened cue selection-using an electrophysiological measure of selective attention. Specifically, we used a reverse correlation approach in combination with frequency-tagged steady-state visual potentials (ssVEPs). Twenty-one participants completed a modified dot-probe task in which each member of a pair of lateralized face cues, varying in emotional expression (angry-angry, neutral-angry, neutral-neutral), flickered at one of two frequencies (15 or 20 Hz), to evoke ssVEPs. One cue was then replaced by a response probe, and participants indicated the probe orientation (0° or 90°). We analyzed the ssVEP evoked by the cues as a function of response speed to the subsequent probe (i.e., a reverse correlation analysis). Electrophysiological measures of cue processing varied with probe hemifield location: Faster responses to left probes were associated with weak amplification of the preceding left cue, apparent only in a median split analysis. By contrast, faster responses to right probes were systematically and parametrically predicted by diminished visuocortical selection of the preceding right cue. Together, these findings highlight the poor validity of the dot-probe task, in terms of quantifying intrinsic, nondirected attentive selection irrespective of probe/cue location.

Responding to emotional scenes: effects of response outcome and picture repetition on reaction times and the late positive potential.

Processing the motivational relevance of a visual scene and reacting accordingly is crucial for survival. Previous work suggests the emotional content of naturalistic scenes affects response speed, such that unpleasant content slows responses whereas pleasant content accelerates responses. It is unclear whether these effects reflect motor-cognitive processes, such as attentional orienting, or vary with the function/outcome of the motor response itself. Four experiments manipulated participants' ability to terminate the picture (offset control) and, thereby, the response's function and motivational value. Attentive orienting was manipulated via picture repetition, which diminishes orienting. A total of N = 81 participants completed versions of a go/no-go task, discriminating between distorted versus intact pictures drawn from six content categories varying in positive, negative, or neutral valence. While all participants responded faster with repetition, only participants without offset control exhibited slower responses to unpleasant and accelerated responses to pleasant content. Emotional engagement, measured by the late positive potential, was not modulated by attentional orienting (repetition), suggesting that the interaction between repetition and offset control is not due to altered emotional engagement. Together, results suggest that response time changes as a function of emotional content and sensitivity to attention orienting depends on the motivational function of the motor response.

The malleability of emotional perception: Short-term plasticity in retinotopic neurons accompanies the formation of perceptual biases to threat.

Emotional experience changes visual perception, leading to the prioritization of sensory information associated with threats and opportunities. These emotional biases have been extensively studied by basic and clinical scientists, but their underlying mechanism is not known. The present study combined measures of brain-electric activity and autonomic physiology to establish how threat biases emerge in human observers. Participants viewed stimuli designed to differentially challenge known properties of different neuronal populations along the visual pathway: location, eye, and orientation specificity. Biases were induced using aversive conditioning with only 1 combination of eye, orientation, and location predicting a noxious loud noise and replicated in a separate group of participants. Selective heart rate-orienting responses for the conditioned threat stimulus indicated bias formation. Retinotopic visual brain responses were persistently and selectively enhanced after massive aversive learning for only the threat stimulus and dissipated after extinction training. These changes were location-, eye-, and orientation-specific, supporting the hypothesis that short-term plasticity in primary visual neurons mediates the formation of perceptual biases to threat. (PsycINFO Database Record

Assessing the internal consistency of the event-related potential: An example analysis.

ERPs are widely and increasingly used to address questions in psychophysiological research. As discussed in this special issue, a renewed focus on questions of reliability and stability marks the need for intuitive, quantitative descriptors that allow researchers to communicate the robustness of ERP measures used in a given study. This report argues that well-established indices of internal consistency and effect size meet this need and can be easily extracted from most ERP datasets, as demonstrated with example analyses using a representative dataset from a feature-based visual selective attention task. We demonstrate how to measure the internal consistency of three aspects commonly considered in ERP studies: voltage measurements for specific time ranges at selected sensors, voltage dynamics across all time points of the ERP waveform, and the distribution of voltages across the scalp. We illustrate methods for quantifying the robustness of experimental condition differences, by calculating effect size for different indices derived from the ERP. The number of trials contributing to the ERP waveform was manipulated to examine the relationship between signal-to-noise ratio (SNR), internal consistency, and effect size. In the present example dataset, satisfactory consistency (Cronbach's alpha > 0.7) of individual voltage measurements was reached at lower trial counts than were required to reach satisfactory effect sizes for differences between experimental conditions. Comparing different metrics of robustness, we conclude that the internal consistency and effect size of ERP findings greatly depend on the quantification strategy, the comparisons and analyses performed, and the SNR.