Benjamin T. Vincent

The prominence of behavioural biases in eye guidance

When attempting to understand where people look during scene perception, researchers typically focus on the relative contributions of low- and high-level cues. Computational models of the contribution of low-level features to fixation selection, with modifications to incorporate top-down sources of information have been abundant in recent research. However, we are still some way from a model that can explain many of the complexities of eye movement behaviour. Here we show that understanding biases in how we move the eyes can provide powerful new insights into the decision about where to look in complex scenes. A model based solely on these biases and therefore blind to current visual information outperformed popular salience-based approaches. Our data show that incorporating an understanding of oculomotor behavioural biases into models of eye guidance is likely to significantly improve our understanding of where we choose to fixate in natural scenes.

The long and the short of it: Spatial statistics at fixation vary with saccade amplitude and task

We recorded over 90,000 saccades while observers viewed a diverse collection of natural images and measured low level visual features at fixation. The features that discriminated between where observers fixated and where they did not varied considerably with task, and the length of the preceding saccade. Short saccades (<8 degrees) are image feature dependent, long are less so. For free viewing, short saccades target high frequency information, long saccades are scale-invariant. When searching for luminance targets, saccades of all lengths are scale-invariant. We argue that models of saccade behaviour must account not only for task but also for saccade length and that long and short saccades are targeted differently.

Do we look at lights? Using mixture modelling to distinguish between low- and high-level factors in natural image viewing

The allocation of overt visual attention while viewing photographs of natural scenes is commonly thought to involve both bottom-up feature cues, such as luminance contrast, and top-down factors such as behavioural relevance and scene understanding. Profiting from the fact that light sources are highly visible but uninformative in visual scenes, we develop a mixture model approach that estimates the relative contribution of various low and high-level factors to patterns of eye movements whilst viewing natural scenes containing light sources. Low-level salience accounts predicted fixations at luminance contrast and at lights, whereas these factors played only a minor role in the observed human fixations. Conversely, human data were mostly explicable in terms of a central bias and a foreground preference. Moreover, observers were more likely to look near lights rather than directly at them, an effect that cannot be explained by low-level stimulus factors such as luminance or contrast. These and other results support the idea that the visual system neglects highly visible cues in favour of less visible object information. Mixture modelling might be a good way forward in understanding visual scene exploration, since it makes it possible to measure the extent that low-level or high-level cues act as drivers of eye movements.

Optimal feature integration in visual search

Despite embodying fundamentally different assumptions about attentional allocation, a wide range of popular models of attention include a max-of-outputs mechanism for selection. Within these models, attention is directed to the items with the most extreme-value along a perceptual dimension via, for example, a winner-take-all mechanism. From the detection theoretic approach, this MAX-observer can be optimal under specific situations, however in distracter heterogeneity manipulations or in natural visual scenes this is not always the case. We derive a Bayesian maximum a posteriori (MAP)-observer, which is optimal in both these situations. While it retains a form of the max-of-outputs mechanism, it is based on the maximum a posterior probability dimension, instead of a perceptual dimension. To test this model we investigated human visual search performance using a yes/no procedure while adding external orientation uncertainty to distracter elements. The results are much better fitted by the predictions of a MAP observer than a MAX observer. We conclude a max-like mechanism may well underlie the allocation of visual attention, but this is based upon a probability dimension, not a perceptual dimension.

Is the early visual system optimised to be energy efficient

This paper demonstrates that a representation which balances natural image encoding with metabolic energy efficiency shows many similarities to the neural organisation observed in the early visual system. A simple linear model was constructed that learned receptive fields by optimally balancing information coding with metabolic expense for an entire visual field in a 2-stage visual system. The input to the model consists of a space variant retinal array of photoreceptors. Natural images were then encoded through a bottleneck such as the retinal ganglion cells that form the optic nerve. The natural images represented by the activity of retinal ganglion cells were then encoded by many more ‘cortical’ cells in a divergent representation. Qualitatively, the system learnt by optimising information coding and energy expenditure and matched (1) the centre surround organisation of retinal ganglion cells; (2) the Gabor-like organisation of cortical simple cells; (3) higher densities of receptive fields in the fovea decreasing in the periphery; (4) smaller receptive fields in the fovea increasing in size in the periphery; (5) spacing ratios of retinal cells; and (6) aspect ratios of cortical receptive fields. Quantitatively, however, there are small but significant discrepancies between density slopes which may be accounted for by taking optic blur and fixation induced image statistics into account. In addition, the model cortical receptive fields are more broadly tuned than biological cortical neurons; this may be accounted for by the computational limitation of modelling a relatively low number of neurons. This paper shows that retinal receptive field properties can be understood in terms of balancing coding with synaptic energy expenditure and cortical receptive fields with firing rate energy expenditure, and provides a sound biological explanation of why ‘sparse’ distributions are beneficial.

Synaptic energy efficiency in retinal processing

Recent work suggests that the visual system may represent early visual information in an energy efficient manner [Nature 381 (1996); Nature, 381 (1996) 607; Neural Comput. 3 (2001) 799; Curr. Opin. Neurobiol. 11 (2001) 475]. This paper applies the idea of energy efficient representations to understand retinal processing, and provides evidence that centre surround processing observed is efficient in terms of minimizing synaptic activity. In particular, it is shown that receptive fields at different retinal eccentricities and at different levels of noise, can be understood in terms of maximizing the transmission of visual information given a constraint on total synaptic strengths and hence energy consumption.

Covert visual search: Prior beliefs are optimally combined with sensory evidence

Has evolution optimized visual selective attention to make the best possible use of all information available? If so, then Bayesian optimal performance in a localization task is achieved by optimally weighting the visual evidence with ones prior spatial expectations. In 2 psychophysical experiments, participants conducted covert target localization where both visual cues and prior expectations were available. The amount of information conveyed by the visual evidence was held constant, while the degree of belief was manipulated via peripheral cuing (Experiment 1) and spatial probabilities (Experiment 2). A number of findings result: (1) People appear to optimally combine slightly biased prior beliefs with sensory evidence. (2) These biases are directly comparable to those descriptively accounted for by the Prospect Theory. (3) Probabilistic information about a targets upcoming location is integrated identically, irrespective of whether endogenous or exogenous cuing is used. (4) In localization tasks, spatial attention can be understood and quantitatively modeled as a set of prior expectations over space that modulate incoming noisy sensory evidence.

Quantitative Analysis of Three-Dimensional Human Mammary Epithelial Tissue Architecture Reveals a Role for Tenascin-C in Regulating c-Met Function

Remodeling of the stromal extracellular matrix and elevated expression of specific proto-oncogenes within the adjacent epithelium represent cardinal features of breast cancer, yet how these events become integrated is not fully understood. To address this question, we focused on tenascin-C (TN-C), a stromal extracellular matrix glycoprotein whose expression increases with disease severity. Initially, nonmalignant human mammary epithelial cells (MCF-10A) were cultured within a reconstituted basement membrane (BM) where they formed three-dimensional (3-D) polarized, growth-attenuated, multicellular acini, enveloped by a continuous endogenous BM. In the presence of TN-C, however, acini failed to generate a normal BM, and net epithelial cell proliferation increased. To quantify how TN-C alters 3-D tissue architecture and function, we developed a computational image analysis algorithm, which showed that although TN-C disrupted acinar surface structure, it had no effect on their volume. Thus, TN-C promoted epithelial cell proliferation leading to luminal filling, a process that we hypothesized involved c-met, a proto-oncogene amplified in breast tumors that promotes intraluminal filling. Indeed, TN-C increased epithelial c-met expression and promoted luminal filling, whereas blockade of c-met function reversed this phenotype, resulting in normal BM deposition, proper lumen formation, and decreased cell proliferation. Collectively, these studies, combining a novel quantitative image analysis tool with 3-D organotypic cultures, demonstrate that stromal changes associated with breast cancer can control proto-oncogene function.

Search asymmetries: parallel processing of uncertain sensory information.

What is the mechanism underlying search phenomena such as search asymmetry? Two-stage models such as Feature Integration Theory and Guided Search propose parallel pre-attentive processing followed by serial post-attentive processing. They claim search asymmetry effects are indicative of finding pairs of features, one processed in parallel, the other in serial. An alternative proposal is that a 1-stage parallel process is responsible, and search asymmetries occur when one stimulus has greater internal uncertainty associated with it than another. While the latter account is simpler, only a few studies have set out to empirically test its quantitative predictions, and many researchers still subscribe to the 2-stage account. This paper examines three separate parallel models (Bayesian optimal observer, max rule, and a heuristic decision rule). All three parallel models can account for search asymmetry effects and I conclude that either people can optimally utilise the uncertain sensory data available to them, or are able to select heuristic decision rules which approximate optimal performance.

Bayesian accounts of covert selective attention: A tutorial review

Decision making and optimal observer models offer an important theoretical approach to the study of covert selective attention. While their probabilistic formulation allows quantitative comparison to human performance, the models can be complex and their insights are not always immediately apparent. Part 1 establishes the theoretical appeal of the Bayesian approach, and introduces the way in which probabilistic approaches can be applied to covert search paradigms. Part 2 presents novel formulations of Bayesian models of 4 important covert attention paradigms, illustrating optimal observer predictions over a range of experimental manipulations. Graphical model notation is used to present models in an accessible way and Supplementary Code is provided to help bridge the gap between model theory and practical implementation. Part 3 reviews a large body of empirical and modelling evidence showing that many experimental phenomena in the domain of covert selective attention are a set of by-products. These effects emerge as the result of observers conducting Bayesian inference with noisy sensory observations, prior expectations, and knowledge of the generative structure of the stimulus environment.