iiinova — where your digital self keeps evolving

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

And when presenting distractors during the delay? Can these models hold on to the target and ignore the distractor? Again: 🚫 STSP model, ✅ Feedback model. 10/n

120

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

When cued to remember just one of two targets, can these models flexibly select the cued one? 🚫 STSP model, ✅ Feedback model. You may think “Pshh, you cheater, the cue cannot even reach EVC”, but even alternative cueing strategies did not allow the STSP model to be flexible. 9/n

120

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

Can we achieve the same with a Feedback mechanism? The short answer is, yes! This is cool, it shows that both models produce latent visual thoughts in EVC. We’re ready to throw some real tasks at these things now! 8/n

110

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

BUT FIRST… Can we use our neural network models to produce latent states in Early Visual Cortex? Others have shown this before using STSP, and we indeed replicate this! Notice the transient EVC dynamics in response to the target, and recovered selectivity in response to the impulse. 7/n

110

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

For the STSP model, we added an STSP mechanism on the Early Visual Cortex module: Neurotransmitter resources deplete rapidly after firing, but increased calcium keeps neurons potentiated. For the Feedback model, we added feedback connections from the frontal to the EVC module. 5/n

110

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

The neural network models: Have an EVC & frontal module, each with a left and right hemisphere. A hemisphere is a circuit of orientation selective neural populations. Circuits have identical architecture, but are tuned to produce transient (EVC) or ring attractor (frontal) dynamics. 4/n

110

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

The 2nd account is Feedback: Sustained activity in higher-order brain areas provide feedback to sensory areas, which is how visual thoughts end up in Early Visual Cortex. Based on prior findings, we hypothesize that also Feedback results in sub-threshold / latent states in EVC. 3/n

110

Rosanne Rademaker @rademaker.bsky.social · 19/05/2026

The 1st account is Short Term Synaptic Potentiation (STSP): When a target input activates target-selective neurons, those neurons become potentiated. A neutral impulse is more likely to activate previously potentiated neurons, recovering the latent visual thoughts / visual memory. 2/n

300

Rosanne Rademaker @rademaker.bsky.social · 21/01/2026

No matter the exact time point, no matter how we quantified the shift, no matter if we looked at decoding or at representational geometry ¬– the reference frame used by the brain to represent orientations was always smack dab in between retinocentric and allocentric 9/n

350

Rosanne Rademaker @rademaker.bsky.social · 21/01/2026

Well, throughout perception (when the orientation is on the screen) as well as the entire memory delay (the orientation is held in mind), we discovered a reference frame that is in between retinocentric and allocentric coordinates! 8/n

150

Rosanne Rademaker @rademaker.bsky.social · 21/01/2026

Now, even if the pattern *completely shifts* with head tilt, standard (within time point) decoding can only ever infer the exact same label! After all, we as researchers do not know the underlying shift, only the orientation (and hence the label) that was on the screen. 5/n

120

Rosanne Rademaker @rademaker.bsky.social · 21/01/2026

We want to decode visual orientation from the EEG signal to uncover the reference frame used by the brain. But we have a problem… A decoder only learns the association between a label (e.g., 45º) and a pattern of brain activity. Presented with a new pattern of activity, the label is inferred. 4/n

120

Rosanne Rademaker @rademaker.bsky.social · 21/01/2026

Do visual parts of the brain represent visual information in an allocentric or retinocentric reference frame? We used a simple orientation recall task while measuring electroencephalography (EEG) signals from human visual cortex. People had their head upright 😀 or tilted 🫠! 3/n

120

Rosanne Rademaker @rademaker.bsky.social · 21/01/2026

Visual information in our environment is anchored to an allocentric reference frame – a tall building remains upright even when you tilt your head. But head tilt changes the retinal projection of the building from vertical to diagonal. The building is diagonal in a retinocentic reference frame. 2/n

130

Rosanne Rademaker @rademaker.bsky.social · 17/12/2025

We looked hysteresis effects (yes they exist in these data!), the role of eye movements (no they can't explain these findings), and more. But importantly, people are MUCH BETTER at recalling the speed of a single dot moving around fixation, then the speed of more texture-like dot motion!! 8/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/12/2025

Another cool finding: The memory target and the probe could either move in congruent (e.g., both clockwise) or incongruent (e.g., target moved clockwise, the probe counterclockwise) directions. Speed recall was better for congruent motion! 7/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/12/2025

The stimulus was presented for 4–6 seconds, and remembered for 1–8 seconds. This mattered not for dot motion (blue), but it *did* matter for the single dot (red) such that errors were lower when people had more time to encode the speed, and higher at longer delays. 6/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/12/2025

Replicating previous findings, we find that responses become less precise with faster speeds (i.e., response distributions become wider). Participants have a tendency to overestimate the speed of dot motion – recalling it faster than it actually was. 5/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/12/2025

High time for a face off between types of motion. People were shown dot motion (blue) or a single dot (red) rotating around fixation at 6 different speeds. Speed was recalled by adjusting a probe to match the memory. We manipulated a bunch of other things, but more on that later. 4/n

100

Rosanne Rademaker @rademaker.bsky.social · 02/07/2025

Bonus re-analysis of a whole bunch of EEG data: Spatial position decoding is SO MUCH better along the horizontal compared to the vertical meridian! N/N

020

Rosanne Rademaker @rademaker.bsky.social · 02/07/2025

We replicated the horizontal-vertical decoding differences for centrally presented gratings using independent data. But look! This decoding difference is entirely absent for laterally presented gratings! 12/N

110

Rosanne Rademaker @rademaker.bsky.social · 02/07/2025

And remember vignetting? Higher orientation energy at grating-edges aligned with an orientation contributes to decoding (Roth et al., 2018). Visual field anisotropies and vignetting can interact. 10/N elifesciences.org/articles/37241

110

Rosanne Rademaker @rademaker.bsky.social · 02/07/2025

But there’s an inverse problem here: The underlying neural tuning cannot be inferred due to model mimicry: Many possible “sensory coding schemes” can generate very similar outcomes. Look what happens when we consider a slightly larger parameter space… 7/N

120

Rosanne Rademaker @rademaker.bsky.social · 02/07/2025

This sensory coding scheme embeds a prior (from natural image statistics) via redistribution of orientation tuning functions, with more functions around horizontal vs. vertical – not predicted by previous proposals! But… How to recover sensory tuning from EEG data? 5/N

120

Rosanne Rademaker @rademaker.bsky.social · 16/05/2025

Welcome #VSS2025! For those attending the sunny, beaches, and science at this years Florida conference, make sure not to miss the awesome talks and posters from our lab!

0143

Rosanne Rademaker @rademaker.bsky.social · 17/03/2025

So... does visual cortex represent orientations veridically or categorically? During perception, the veridical model does a generally better job. During short-term memory, representational geometry becomes increasingly more categorical along the visual hierarchy!

160

Rosanne Rademaker @rademaker.bsky.social · 17/03/2025

We model what similarity of fMRI responses might look like when assuming that (1) the brain uses a veridical way to represent orientations, or (2) the brain categorizes orientations based on how different they appear in a more experiential sense.

120

Rosanne Rademaker @rademaker.bsky.social · 17/03/2025

Girshick, Landy, & Simoncelli (2011) showed it best: The natural world has more horizontal and vertical ("cardinal") orientations in it than those in between ("obliques"). We start from this uneven probability distribution over orientation.

120

Rosanne Rademaker @rademaker.bsky.social · 17/03/2025

To answer how perception and memory differ in visual cortex, we use line-images ("gratings") with different orientations that people look at, or remember, while we record voxel responses. We correlate patterns of responses evoked by all possible orientations & VOILA, "representational geometry"

120

Rosanne Rademaker @rademaker.bsky.social · 17/03/2025

Early visual cortex processes what we see around us, but also has information about images briefly held in mind. The two must be different... But different how?

120

Rosanne Rademaker @rademaker.bsky.social · 17/09/2024

And what about the distractor itself, which is directly perceived throughout the delay? We show that attention actually enhances these sensory representations, especially when distractor orientation is attended. 7/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/09/2024

So, does attention negatively impact decoding of the orientation held in memory? Yes it does! Throughout the visual hierarchy we see a hit to decoding of the remembered target when attention is paid to the distractor 6/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/09/2024

Recall of the remembered orientation was worse when attention was diverted towards changes in distractor contrast and orientation. Univariate BOLD was also higher with attention, compared to when the distractor was ignored. 5/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/09/2024

We scanned people over the course of several days. While they remembered an orientation, we showed another stimulus on the screen. This distractor had several small changes in its contrast or orientation on every trial. We manipulated just one thing: Attention. 4/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/09/2024

But everyday life also requires attention to the environment, so you can navigate it safely. When you hold an image in mind, this often happens concurrently with attention to visual inputs! Can visual cortex concurrently process competing top-down demands? 3/n

100

Rosanne Rademaker @rademaker.bsky.social · 17/09/2024

Everyday life typically requires visual memory when there are concurrent visual inputs. Together with John Serences, we found out a while back that we can use fMRI to decode the contents of visual working memory from early sensory cortex, even during such concurrent visual input. 2/n

110