iiinova — where your digital self keeps evolving

Laurent Caplette @laucaplette.bsky.social · 28/08/2026

I'm excited to share that I'll be joining Université de Montréal's School of Optometry as an Assistant Professor this fall! My lab will focus on understanding how we process and represent the dynamic and complex visual world, using AI, behavior and neuroimaging. Recruiting grad students for 2027!

2153

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

In the end, we recommend: 1) stopping the use of the existing CV Euclidean distance (except in very specific cases) and using the GCV or WCC variant 2) stopping the use of the WCC and existing CV correlation distance, and use our GCV variant. See preprint for details and nuances: t.ly/XaePh (13/13)

t.ly

Improved cross-validated distances for multivariate pattern analysis

Characterizing the dissimilarity of neural representations between experimental conditions, and tracking it across time, is a central goal of multivariate pattern analysis. Guggenmos et al. (2018) ass...

010

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

Finally, we compare this "generalized cross-validation" (GCV; patterns averaged before computing distances) to "within-class correction" (WCC; distances computed and then averaged). Because of the bias intrinsic to correlations (noise => r tends to 0), the GCV correlation distance is more accurate.

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

Although the previous formulation has slightly increased reliability (at onset) at first sight (pabel b), this is artificially created by enforcing a clipping between 0 and 2. When clipping is removed, our new formulation is more reliable (panel d). (11/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

This formulation is more accurate than the previous one. It also has a more expected time course. (10/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

When we generalize in the same way we did for the Euclidean distance, we get: (9/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

We can make use of that and propose another way to cross-validate the correlation distance: (8/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

But the correlation distance is related to the Euclidean distance of Z-transformed vectors: (7/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

For the correlation distance, cross-validation cannot make the distance fully unbiased. However, when done correctly it can have some benefits. Guggenmos et al (2018) proposed an equation (below). However, it needs arbitrary regularization, otherwise estimates could explode or be imaginary. (6/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

Turns out, when we do this minor adjustment, reliability (across MEG sessions) and accuracy (to a simulated ground truth) are slightly improved (the new distance is the GCV in pink). (5/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

We can make use of the fact that condition order is usually random, and therefore partitions are arbitrary, to include within-partition distances in that equation, while keeping it unbiased. (This gets us closer to within-class correction; a word on that later.) (4/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

First, we show that the CV Euclidean distance is equivalent to a sum of between-partition distances. (3/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

The basic Euclidean distance is biased: it increases with noise. This limits the analyses that can be done with it. The CV Euclidean distance (Walther et al., 2016) solves this bias by using patterns from two independent data partitions (see below: conditions x and y and partitions A and B). (2/13)

100

Laurent Caplette @laucaplette.bsky.social · 14/08/2026

Another new preprint! t.ly/XaePh I had some math fun with distance equations and found out some stuff 🤓 We propose cross-validated (CV) Euclidean + correlation distances with improved reliability and accuracy, for RSA and beyond! 🧵👇(1/13) ⚠️Many equations below 📈but some MEG data as well!

t.ly

Improved cross-validated distances for multivariate pattern analysis

150

Laurent Caplette @laucaplette.bsky.social · 12/08/2026

New preprint! t.ly/yxfsh Several EEG features are commonly analyzed to characterize brain development. We showed that the latent structure of these features (their distances) also differed between children and adults, and that these distances are much more robust to non-developmental factors.

020

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

In conclusion, we believe that this approach holds promise for characterizing perception in greater detail. More generally, we hope that our work helps neuroscientists and psychologists recognize the important distinctions between different "types" of time! (7/7) t.ly/9DgIF

t.ly

Time^2: A framework for the neural dynamics of visual perception

Whenever we look at an object, we seem to perceive it immediately. However, this is not the case for two reasons. First, it takes hundreds of milliseconds for the brain to process visual information r...

000

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

When considering both temporal dimensions simultaneously, we can also disentangle between competing explanations of neural latencies in high-level areas: Was feature B processed more slowly than A or was it simply attended later? (6/7)

100

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

We showed that rhythmic sampling, rather than rhythmic processing, is the most prevalent consequence of ongoing brain oscillations. Our method should also help characterize neural mechanisms of predictions, accumulation, adaptation... (5/7)

110

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

We argue that it is essential to consider both at the same time to obtain a more complete portrait of visual perception and disentangle between competing explanations of neural phenomena. In a previous paper (Caplette et al., 2023, J Neurosci), we developed an approach to do so: Time^2. (4/7)

100

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

These two temporal dimensions are often conflated in the literature. And most studies have considered only one of them. Some examples: (3/7)

100

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

When we look at a visual scene, our brain has to process the visual information and this takes some time. During that time, visual information is also continuously entering the visual system (even if the stimulus is static). We call these two temporal facets processing time and stimulus time. (2/7)

100

Laurent Caplette @laucaplette.bsky.social · 06/08/2026

Our Viewpoint paper just got accepted at Journal of Neuroscience! @sfnjournals.bsky.social We show that there are actually *two* temporal "dimensions" to consider in visual perception and we discuss the promises of a method to disentangle them. 🧵👇(1/7) t.ly/9DgIF

181

Laurent Caplette @laucaplette.bsky.social · 14/06/2026

Come see our poster at OHBM! We looked at how various neural features are organized in children vs adults. Turns out not only feature values differ across development, also their organization! Moreover, such feature relationships are way more robust to task/stimuli than feature values.

Two MDS plots representing how multiple neural features are inter-related, in children and in adults

010

Laurent Caplette @laucaplette.bsky.social · 21/03/2026

After a mental health hiatus and trying something else for a while, I’m back (on here and in science)! 😄 Couple updates coming in the next months. It seems like a lot more people have arrived on bluesky in the last 18 months… Excited to see what everyone’s been up to! 🙂

050

Laurent Caplette @laucaplette.bsky.social · 16/05/2024

It's #VSS! Come see our poster Saturday PM about the representations of expectations across time and brain regions! #VSS2024

020

Reposted by Laurent Caplette

paolo @paolopalma.bsky.social · 13/05/2024

I don’t know if people here have seen this here but there is a new #psychjob wiki sites.google.com/view/psychjo... #SocialPsych #Psychology #PsySciSky

Screenshot of LinkedIn post by Amédee Marchand Martella: The Psych Job Wiki (a fantastic resource for finding positions in academia for psychology & educational psychology) is ending and as such, Alyssa Lawson and I have created a new site to keep the wiki going. For those of you (or anyone you know) who will be on the job market this coming academic year, looking for tenure track, non-tenure track, or postdoc positions, here is a link to the new wiki. Feel free to spread the word, use the wiki, and contribute.

05049

Laurent Caplette @laucaplette.bsky.social · 06/12/2023

VSS is down at abstract submission time 😱

020

Laurent Caplette @laucaplette.bsky.social · 06/12/2023

Also experiencing this here in the US... 😬

100