• No results found

The aim of this thesis was twofold: 1) to establish neural mechanisms for widely observed perceptual deficits induced by high perceptual load tasks on

perceptual systems, and 2) to create a predictive model of the perceptual demands of a visual task using the visual information available in the task itself. In Chapters 3 and 4, recent neuroimaging methods were employed to measure the neural excitation associated with oriented visual stimuli while participants completed tasks of both low and high perceptual load. Chapter 5 expanded and developed modelling approaches rooted in computer vision and machine

learning to estimate perceptual load in the dynamic real-world task of driving.

In Experiment 1 of Chapter 3, deleterious effects of high perceptual load were observed on orientation change detection using novel experimental stimuli and psychophysical design, consistent with previous findings of reduced stimulus detection under high load (e.g. Macdonald & Lavie, 2008; Carmel et al, 2007). While perceptual load was modulated with a primary RSVP task at fixation, orientation change detection accuracies were measured across ranges of orientation change offsets in a secondary peripheral task. It was found that high load led to an overall suppression of change detection accuracy across the offset range and an increased orientation offset threshold for detection, thus confirming perceptual load effects on the perception of oriented grating stimuli. In Experiment 2, multivariate pattern analysis (MVPA) methods were used to quantify the representational content of visual areas V1, V2, and V3, in

response to large oriented gratings under low and high perceptual load. While univariate BOLD signal analysis confirmed a general suppression of activity in

160 cortical areas, no statistical difference was found by MVPA methods between pattern classification accuracy for orientation-specific activity patterns elicited under low and high load conditions. This unexpected result is perhaps best explained by properties of the experimental design: specifically, participants' spare attentional capacity under low load may not have been allocated to perceiving the orientation of the peripheral grating due to attentional capture by other properties of the stimulus, and furthermore altering the range of oriented gratings to enable the construction of voxel tuning function (VTFs), which characterise the tuning properties of neural populations explicitly.

The design of the experiment in Chapter 4 was therefore modified to address these possibilities. A secondary delayed orientation discrimination task was included to ensure spill-over of resources to the perception of orientation, while the orientation range of the grating stimuli was extended across the full range to allow the construction of voxel tuning functions (VTFs). Perceptual load was again manipulated with a central primary task, performance on which was reduced in the high load condition indicating the effectiveness of the load

manipulation. Performance on the secondary orientation discrimination task was high under both load conditions (more than 80% detection accuracy for 20° offsets), suggesting that perceptual resources were indeed directed to the perception of the gratings, and furthermore performance was decreased when the primary task load was high. Analysis of BOLD signal again found a general suppression of visuo-cortical activity across early visual cortex in the high load condition, replicating previous findings of reduced visual processing under load (Rees et al., 1997; Schwartz et al., 2005).

To investigate orientation-specific modulations, VTFs were then constructed using responses elicited by the gratings across V1, V2, and V3: while no statistical difference was found between VTFs extracted for V2 and V3 voxels,

161 tuning functions constructed from V1 activity were found to have both reduced amplitude and increased bandwidth under high perceptual load relative to low. This finding suggests a novel mechanism of action for load-induced perceptual deficits which originates at the very earliest stages of cortical processing.

Interestingly, SVM-based MVPA using distributed patterns of activity in Chpater 4 also showed no significant difference for orientation classification accuracy between activity elicited under high and low load in any of the investigated visual areas (V1, V2, and V3). This apparent discrepancy between VTF and MVPA findings may be resolved by the fact that MVPA is a representational measure across sets of voxels whereas VTF analysis characterises the tuning properties of neural populations within individual voxels. It does not follow then that MVPA accuracy is always positively correlated with VTF multiplicative or bandwidth scaling, for example in the case where the distribution of voxel orientation preferences is non-uniform. This possibility was addressed, where it was found that orientation preferences across voxels were not uniform across the probed orientation range, voxel preferences being biased towards the horizontal meridian. The finding suggests that perceptual load acts to degrade orientaion encoding at the level of local neural populations in cortex rather than across whole retinotopic areas.

In Chapter 5 a predictive model of perceptual load was produced using computer vision and machine learning techniques. A dataset of 1809 short video clips was collected depicting real-world driving and a value of perceptual load assigned to each clip through aggregating more than 60,000 pairwise comparisons between clips, collected in a large-scale experiment. Spatio- temporal features, used successfully in previous work to classify motion-based actions from video (i.e. improved dense trajectories; Wang et al., 2013; 2015; and convolutional 3D features; Du Tran et al, 2015), were extracted from the

162 video clips to produce parsimonious clip representations. IDT and C3D

representations were then fused using a multichannel kernel and mapped to the perceptual load values derived from pairwise comparisons using regression analyses. A variety of methods for feature fusion and regression were

compared: it was found that non-linear feature fusion, utilising a combination of all IDT and C3D features, produced improved regression performance over linear methods, indicating that complex interactions between motion features across the visual field are informative for the prediction of perceptual load. The best performing model configuration resulted in 63.7% of variance in perceptual load values being explained by the model. Furthermore, an investigation into the relative importance of certain features in the model's accuracy found that trajectory motions in the x-axis (i.e. horizontal across the visual field) were relatively more useful for the model, confirming intuitions regarding such motions during driving (e.g. pedestrians crossing the road). The system therefore constitutes the first model to predict, a priori from visual information, the perceptual load induced by a complex, dynamic, real-world task.

Related documents