Will Harrison

@willjharrison.bsky.social

949 Followers 409 Following

Vision scientists and horror film enthusiast. I can’t read or send DMs because Australia banned social media accounts for people under 16…. You do the maths. I also lecture at the University of the Sunshine Coast when I'm not at the beach.

Figure 3 | Visual connectivity asymmetries. a. Meridian asymmetries in early visual white-matter connectivity. Left, surface-based representations of the horizontal meridian (HM) and vertical meridian subdivisions, including the lower vertical meridian (LVM) and upper vertical meridian (UVM), in V1. Middle, bar plots showing mean streamline density (Sd) across the four principal meridians (HM, vertical meridian (VM), LVM, and UVM). Right, polar-angle–resolved connectivity profiles between V1 and V2, demonstrating higher connectivity along the horizontal compared with the vertical meridian and along the lower compared with the upper vertical meridian. Results were derived from diffusion MRI tractography using the Human Connectome Project (HCP) dataset (n = 1,061 participants). b. Replication across independent cohorts. Meridian asymmetry profiles derived from the Cambridge Centre for Ageing and Neuroscience (Cam-CAN, n = 584) and Pediatric Imaging, Neurocognition, and Genetics (PING, n = 106) datasets reproduce the horizontal–vertical asymmetry (HVA) and vertical meridian asymmetry (VMA) patterns. c. Meridian asymmetries are absent in long-range visual connectivity. Polar-angle–resolved streamline density profiles for connections between the rest of the brain and V1 show reduced or absent asymmetries, indicating that meridian-dependent connectivity differences arise primarily within the visual system. d. Lifespan trajectories of meridian asymmetries. Individual HVA (left) and VMA (right) indices plotted as a function of age across PING, HCP, and Cam-CAN reveal distinct developmental and aging-related trends. For visualization purposes, data points were histogram-matched across datasets.
A ghostly face
Screenshot of job ad: How do Bayesian brains acquire priors? APPLICATIONS CLOSE 20/4/2026 11:55 PM Summary of the Project This project explores how the brain constructs perceptual experience from visual input, focusing on the role of Bayesian models in perception. A key challenge in vision science is understanding how humans interpret complex scenes from the limited information available in retinal images. Modern theories suggest that perception involves probabilistic inference, where the brain integrates sensory signals with prior expectations to make sense of the world. However, the origins and nature of these expectations remain poorly understood. This research aims to advance our understanding of perceptual experience by examining how structured patterns in visual input can inform models of perception. The work spans computational modelling and experimental approaches to uncover principles that explain how visual systems interpret properties such as shape, material, and lighting from images. By addressing fundamental questions about perception, this project will contribute to psychology, neuroscience, and artificial intelligence, offering insights into how biological and artificial systems can learn to interpret complex environments.
Four panels showing how an image of a kiwi fruit can be filtered to understand contrast energy.
V1 receptive fields in 3 dimensions, x, y and t.
Contrast Energy The contrast energy of THINGS images was well described as a linear transform of spatial fre- quency on log-log axes, with a slope of 1.3 (Figure 2A). There was relatively little variation across images, with the simple 1/f model accounting for 97% of variance across all 26,107 images. Contrast energy as a function of orientation, however, was relatively more heterogeneously distributed across images (Figure 2B). Nonetheless, our model of oriented contrast accounts for 17% of the variance in the THINGS images. There are many plausible reasons for the greater con- sistency of energy as a function of spatial frequency relative to orientation. As one example, Torralba and Oliva (2003) showed that such statistics depend on whether image content is natural or not, with relatively less cardinal-oblique contrast bias for natural images. A strength of the THINGS database is its diversity of image content. Furthermore, by design, the images in the THINGS database are digital photos with objects in various compositions and configurations. For example, some photos tightly frame a specific object, and others are framed with the camera lens rotated by some amount. These variations will have little impact on the 1/f spectrum of the images, which simply quantifies the scale invariance of the contrast distributions, but can greatly impact the distribution of oriented contrast (e.g. a picket fence rotated by 45° will have the same 1/f spectrum as the original, but its oriented contrast distribution will be translated). It is somewhat unsurprising, therefore, that oriented contrast is more variable across images. Perhaps more import- ant to the intended use of THINGS images, however, is that there is no reason to expect that these distributions of contrast energy deviate meaningfully from those encountered in natural settings.
A chatGPT prompt saying "create an image of the text "the quick brown fox jumped over the lazy dog", and circle all the vowels in the image." ChatGPT generated an image that says "The quick brown fox jumpeed over the lazy dog" and circled some vowels, but also the dot of an "i", and it missed several vowels.
A comic showing a boy's reaction and a robot's reaction to a bowl of ice cream. The boy says "it's ice cream", while the robot says, "It's a spoon"
Screenshot of an article in Biometrika, 1908, called "The probably error of a mean", by "STUDENT"
A grid of Abraham Lincoln faces, parametrically distorted along different axes.
A celebrity selfie taken by Bradley Cooper, featuring many other people including Ellen DeGeneres. The image is distorted in a way that makes faces appear strange, but leaves non-face information intact.
Fig. 3. Data for Experiment 2 (a) and Experiment 3 (b). For both (a) and (b), at left is shown median rank-order timeline placement of each image in the sequence, as a function of the true index of the image (darker colors represent later true temporal positions), separately for filler images (top) and targets (bottom). In the middle is shown the average timeline placement for fillers and targets. Each dot or line represents a participant. At right is shown the difference in timeline placement (fillers – targets; larger numbers indicate that targets were placed earlier in time) as a function of the number of times a target was repeated (controlling for recognition memory and initial presentation block). Circles indicate mean across participants; error bars indicate ±1 SEM. *p < .05. ***p < .001.
A 3d bar chart
A scientific image showing an original image (grayscale dog), the same image after processing with oriented edge filters, and then the spectral power as a function of orientation.
Security verification, similar to captcha, but the task is to "please click on the object that only appears once". The image includes coloured letters, shapes, and numbers.
Baller waiting at airport lounge.
Data figures showing heart rates of experiment participants at two time points. At the second time point there is highly suspicious ordering of data that is very unlikely to be natural.
Stimulus variance should be a feature, not a bug. If we develop generalisable theories in cognitive neuroscience, we should be able to predict what people will see, think and feel when they are outside the lab. To mitigate the risk of building theories that solely predict how people will respond to abstract spots of light on a computer monitor, we must embrace naturalistic high-variance stimulus sets. I will describe how observer models can generate accurate quantitative predictions for such naturalistic stimuli, thereby (hopefully) improving our understanding of human psychology in the real world.
Daggy dude who is high quality.
Figure 1. Natural image statistics. (A) Orientations of edges in digital photos have systematic biases. Original photo taken by Rafael Forseck and used under the Unsplash Licence. (B) Idealised distribution of orientations across many natural images (Hansen & Essock, 2004; Harrison et al., 2023).
A presentation slide that reads: "Will Harrison. Sunny Psychophysics. What makes horror villains scary? How do we remember colours? How does the brain coordinate eye movements?"