iiinova — where your digital self keeps evolving

Will Harrison @willjharrison.bsky.social · 24/08/2026

By finding the local peaks and troughs of the filtered noise, it's easy to add specular highlights to give the appearance of gloss.

001

Will Harrison @willjharrison.bsky.social · 23/08/2026

Sometimes I just want to try to code pretty shit low pass filter 3D noise (creates the undulation) -> project noise onto a colour wheel (with some other bells and whistles)

1124

Will Harrison @willjharrison.bsky.social · 12/05/2026

Here's what I think of this study.

010

Will Harrison @willjharrison.bsky.social · 24/04/2026

Scientific figures don't have to be realistic or clever, they just have to communicate ideas in a way that's clearer than a passage of text. Love this figure.

061

Will Harrison @willjharrison.bsky.social · 11/04/2026

The proposed model also makes a specific prediction about population-level neural activity: total spike counts for sub-populations tuned to specific features should be heterogeneous. We re-analysed some V1 data from cat brain, which is consistent with this prediction. 5/

110

Will Harrison @willjharrison.bsky.social · 11/04/2026

This figure outlines how this is achievable, and that such a population code can reproduce Bayesian behavioural biases (ie. perceptual reports with systematic errors) without needing an explicit prior term during sensory inference. 4/

110

Will Harrison @willjharrison.bsky.social · 10/04/2026

Just whipped this up to show what I'm thinking.

100

Will Harrison @willjharrison.bsky.social · 09/04/2026

Also, imagine if these spatial bin sizes were proportional to the cortical magnification factor. It looks to me that connectivity may be approximately constant across eccentricity after adjusting for systematic RF size differences in periphery. Bin choice will have flow on effects for higher areas.

100

Will Harrison @willjharrison.bsky.social · 09/04/2026

Oh wow that is super interesting. I think this is where a foveated observer/neurophys model would very useful: one way to interpret this is that the fovea dominates object vision, and therefore higher-level connectivity is mostly determined by foveal processing?

110

Will Harrison @willjharrison.bsky.social · 08/04/2026

The horizontal bias in visual processing is also evident in functional connectivity.

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.

2152

Will Harrison @willjharrison.bsky.social · 24/03/2026

Today at 4pm I’ll be presenting some scary psychophysics results at the University of Queensland, room 215 of Chamberlain. Come join if you’re around. @brisepsi.bsky.social

1141

Will Harrison @willjharrison.bsky.social · 09/03/2026

I'm looking for a skilled PhD student who doesn't want to work too hard but still do great science: unisc-cp.enquire.cloud/round/RND-00...

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.

02111

Will Harrison @willjharrison.bsky.social · 25/02/2026

So @reubenrideaux.bsky.social and I decided to run an image-processing workshop at this year's EPC/APCV. We will be teaching people how to compute the contrast energy of kiwi fruit, I guess. Sign up now: visualneuroscience.auckland.ac.nz/epc-apcv-2026/ @expsyanz.bsky.social

Four panels showing how an image of a kiwi fruit can be filtered to understand contrast energy.

1156

Will Harrison @willjharrison.bsky.social · 08/02/2026

Matlab nerds, I need help with a coding task that has embarrassingly defeated me. I want to turn a 3D array (kernel defined in 3 dimensions) into a cube visualisation... HALP

V1 receptive fields in 3 dimensions, x, y and t.

321

Will Harrison @willjharrison.bsky.social · 21/11/2025

One of the few papers of my own that I enjoy reading is an analysis of a wonderful image database. But I'm a nerd. Harrison, W. J. (2022). Luminance and Contrast of Images in the THINGS Database: Perception, 51(4), 244–262. doi.org/10.1177/0301...

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.

120

Will Harrison @willjharrison.bsky.social · 12/08/2025

wow ChatGPT 5 really working at the level of a PhD

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.

100

Will Harrison @willjharrison.bsky.social · 29/06/2025

Wtf true for my real and digital pianos Maybe this is why some keys are easier to learn than others (I only just mastered b flat minor!)

050

Will Harrison @willjharrison.bsky.social · 13/05/2025

This important question lays bear the inadequacy of current machine vision approaches to understanding human vision... and it reminds me of this hilarious Ted Adelson cartoon.

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"

181

Will Harrison @willjharrison.bsky.social · 08/05/2025

Seems like another version of Bart's anti-Bayesian demo: www.sciencedirect.com/science/arti...

160

Will Harrison @willjharrison.bsky.social · 06/05/2025

In this figure here, the "non illusory triangle" is a triangle with broken sides, no? In terms of pixels, it's as illusory as the kanizsa (unless I'm missing something). The interaction occurs in the bottom right panel: the illusory kanizsa will be stronger there than bottom left.

110

Will Harrison @willjharrison.bsky.social · 04/05/2025

Hey guys I think I found a really important paper. www.jstor.org/stable/2331554

Screenshot of an article in Biometrika, 1908, called "The probably error of a mean", by "STUDENT"

020

Will Harrison @willjharrison.bsky.social · 02/05/2025

I even generated a demo that was on the cover of the journal.

A grid of Abraham Lincoln faces, parametrically distorted along different axes.

230

Will Harrison @willjharrison.bsky.social · 01/05/2025

This is a demo I created for a paper a few years ago showing that a constant image distortion - which simulates subtle pathology either in the retina or cortex - has different perceptual implications depending on e.g. viewing distance. Notice how the clothes appear normal, but some faces do not.

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.

210

Will Harrison @willjharrison.bsky.social · 24/04/2025

I find this memory bias very interesting: people tend to recall familiar memoranda as having first occurred further back in the past than less familiar memoranda.

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.

130

Will Harrison @willjharrison.bsky.social · 17/04/2025

Just skimmed - this looks important AND fun. Can’t wait to read the detail 👏

020

Will Harrison @willjharrison.bsky.social · 11/03/2025

Accidentally created this neato visualisation

010

Will Harrison @willjharrison.bsky.social · 04/03/2025

hah I couldn't resist flicking through to see if they use a signal detection analysis... then I stumbled upon thishttps://onlinelibrary.wiley.com/cms/asset/c52b3ab3-7fea-48d5-bc3e-6e9e57e041fa/jocd16241-fig-0002-m.jpg

110

Will Harrison @willjharrison.bsky.social · 29/01/2025

Turns out we can predict both the perceptual errors AND the associated confidence quite well from the same front end mechanism. The coloured lines in these panels show the model output at the perceptual decision stage and the confidence stage.

100

Will Harrison @willjharrison.bsky.social · 29/01/2025

We could then build a bunch of perceptual/confidence models that have an image-computable front end. Given the image statistics detected in a randomly oriented target, how should the observer respond, and what should their confidence be?

110

Will Harrison @willjharrison.bsky.social · 29/01/2025

From this paradigm, we get a distribution of "perceptual errors" (black data left panel), which is how much the reported upright deviates from the ground truth, as well as confidence ratings that are naturally associated with differently sized perceptual errors (black data right panel).

100

Will Harrison @willjharrison.bsky.social · 29/01/2025

On each trial, a participant is shown a small piece of a natural image, which has been rotated by some random amount. The participant has to rotate this target so it appears "upright". They then tell us if they are confident (or not) in their response.

100

Will Harrison @willjharrison.bsky.social · 29/01/2025

First off, all aspects of this paper were led by the unstoppable Rebecca West, building off former PhD student Dr Emily A-Izzeddin's work, in collaboration with Dave Sewell, none of whom are on bsky. I'd tell you to hire Bec and Emily, but they are awesome enough to already have postdocs.

120

Will Harrison @willjharrison.bsky.social · 29/01/2025

New from my lab: "Priors for natural image statistics inform confidence in perceptual decisions" We used a natural image statistics approach to investigate the computations underlying perceptual confidence. 🧵 Article: www.sciencedirect.com/science/arti...

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.

37223

Will Harrison @willjharrison.bsky.social · 10/12/2024

Bot checkers are really exploiting cognitive strategies.... this took me far too long!

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.

060

Will Harrison @willjharrison.bsky.social · 20/11/2024

I want to laugh at this, but then again I’m not the one getting paid $500k/year so maybe the joke is on me.

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.

020

Will Harrison @willjharrison.bsky.social · 14/11/2024

At this year’s Australian Cognitive Neuroscience Society Conference I’ll be presenting a talk in the Naturalistic Paradigms symposium. I’ll talk about observer models for a variety of tasks, such as visual search and face recognition, as well as how the same principles can be applied to neuro data.

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.

1131

Will Harrison @willjharrison.bsky.social · 02/10/2024

I'm giving a symposium to a large Aussie Cog Neuro audience in a couple of months, and I'm thinking about making the title of my talk "Stop doing experiments and start building theories". Follow me for more hot takes.

120

Will Harrison @willjharrison.bsky.social · 11/02/2024

Amazing PhD candidate, Rebecca West (not on bsky), led this collab - we exploit natural images to understand the computations underlying confidence in perceptual decisions. #VisionScience "Priors for natural image statistics inform confidence in perceptual decisions" www.biorxiv.org/content/10.1...

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).

0113

Will Harrison @willjharrison.bsky.social · 11/02/2024

Today I'll be introducing myself to the incoming Honours students at the University of the Sunshine Coast!

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?"

010