iiinova — where your digital self keeps evolving

Riccardo Fusaroli @fusaroli.eurosky.social · 5h

Working on that (half of last we’s harvest), but the prep for eating them is a hassle, so oil will be the main outcome

Riccardo Fusaroli @fusaroli.eurosky.social · 15/10/2026

The egg was awesome, I have to say, especially when augmented (disinfected?) w this as my grandad did when he hadn’t already left for work

120 2

Riccardo Fusaroli @fusaroli.eurosky.social · 05/10/2026

grant submitted. Great cthulhu, how I hate writing grants, but this would/will be a really fun project. Wish us luck!

2160 16

Riccardo Fusaroli @fusaroli.eurosky.social · 19/09/2026

Paris was amazing, the science, the discussions, the morning runs, the people. But I might have developed a problem. Apparently I have had eggplant 4 evenings and 3 lunches out of 5... a selection

1100 10

Riccardo Fusaroli @fusaroli.eurosky.social · 16/09/2026

Controversial, but this city does have its charms (and charming running places)

010 1

Riccardo Fusaroli @fusaroli.eurosky.social · 15/09/2026

not the worst place for a morning run

030 3

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

Why did conventions stabilize rather than drift? Participants were pretty conservative: language matched sketch maps & prior conceptual habits carried over even when switching from one city layout to the other. 10/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

Participants readily picked up and reproduced the linguistic conventions handed down by previous generations. However, we did not find cumulative elaboration over time. Instead of progressively optimizing across generations, conventions stabilized almost immediately. 9/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

Does matching your environment pay off? Yes! In the grid environment, receiving horizontal directions significantly sped up reaching the target, whereas vertical descriptions slowed people down. Environment-aligned conceptualizations directly improve communicative efficiency. 8/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

The city layout directly steered spatial strategies: • "Manhattan" generated horizontal descriptions (e.g., street sequences, "turn right at the 4th street") and street-grid maps. • "Barcelona" triggered vertical descriptions (e.g., "turn after the blue building") and landmark-heavy maps. 7/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

We could look at how the different sources of information and produced representations fed into each other. That ended up mattering a lot to understand e.g. transmission of descriptive styles vs. evolution of them. 6/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

We used a transmission chain design: * Phase 1: Follow directions written by a previous generation in the same environment. * Phase 2: Find a new target alone, write directions, and draw a sketch map for the next generation. * Phase 3: Switch cityscapes (to test carryover effects to new envs) 4/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

Does languages adapt to socio-ecological environments? We built two contrasting virtual reality cities: * "Manhattan": a regular, grid-like street network with low-salience textures. * "Barcelona": an irregular, winding layout with salient, colorful buildings. 3/

200 2

Riccardo Fusaroli @fusaroli.eurosky.social · 10/09/2026

TL;DR: DAGs + experiments showing that urban forms act as a cognitive-cultural niche. Grid layouts foster metric descriptions, irregular layouts landmark-based ones. These conventions transmit reliably across generations, but stabilize rapidly rather than progressively accumulating. 2/

132 3

Riccardo Fusaroli @fusaroli.eurosky.social · 05/09/2026

One of the things I find most fascinating in mathematical/computational thinking is the borrowing and development of model components from other domains. I need to model how different coders annotate a corpus? It's like occupancy models in ecology, with some twists! 1/

161 6

Riccardo Fusaroli @fusaroli.eurosky.social · 03/09/2026

Darn it, this stuff keeps being hard! Age of entry affects MSEL at entry, but time lapsed from entry does not. So I had to redraw the last two dags, probably not for the last time!

000 0

Riccardo Fusaroli @fusaroli.eurosky.social · 03/09/2026

Now we are ready to consider that all measures are observations of latent constructs, and therefore we should at least map the constructs that might be relevant in the causal mechanisms. e.g. MSEL-EL and MLU are expressions of the same construct. 6/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 03/09/2026

Now, we shouldn't forget that this is a longitudinal design with mismatched age of entry, so we should model age as a combination of age of entry and time lapsed since entry. That will come in useful later on. 5/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 03/09/2026

However, we cannot forget that the autistic and non autistic kids were matched on language (albeit a different measure of language, MSEL-EL and not MLU)! So we need to plot that matching! (and yes, MSEL-EL and MLU are related, we'll get there) 4/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 03/09/2026

DAG 1 is the core: we are interested in an outcome Y, as a function of age (A) and diagnostic group (D). However, since development is more complex than age, we also measure 3 individual skills (MLU and 2 Vinelands) relevant for Y, which are also determined by D and A. 2/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 03/09/2026

Since I'm slightly crazy, I decided to start drawing the DAG(s) of an apparently simple language-matched longitudinal study of language development in autism, and its implications. This is a work in progress, so sharing as a way of better defining my thoughts. Comments welcome! Thread!

292 9

Riccardo Fusaroli @fusaroli.eurosky.social · 01/09/2026

Really cool thesis by @shoakamine.bsky.social justice came in with the mail. Conversational alignment, critical appraisal of manual/automated annotations, causal inference, what’s not to like?

2102 10

Riccardo Fusaroli @fusaroli.eurosky.social · 23/08/2026

My partner doesn’t like miso and soy sauce, my kids don’t like tahini nor kombu. Uzu is tricky to find. So I ended up making an Italian version. parmesan crust broth w oatmilk and ginger, homemade egg-noodles, caramelized tomatoes, shiitake and lightly fried chicken. And lemonade peel. 2/

230 3

Riccardo Fusaroli @fusaroli.eurosky.social · 23/08/2026

How I got to empathise w the german way to Italian cuisine. Last time i was in Tokyo i took a ramen course. Amazing. I then fell in love w the vegan ramen uzu at the Teamlab planet installation. I’ve been working on reproducing it. Yet 1/

160 6

Riccardo Fusaroli @fusaroli.eurosky.social · 14/08/2026

Current vs past dwellers along my morning run.

040 4

Riccardo Fusaroli @fusaroli.eurosky.social · 28/07/2026

The locals’ suggested running route might not be too conductive to running, but it does have its charms

0140 14

Riccardo Fusaroli @fusaroli.eurosky.social · 13/07/2026

the bayesian workflow book is great so far, filled with experience based reflections and suggestions. I can also see the different voices. I'd bet @rmcelreath.bsky.social would have written at least 10 pages for the early 1p list of useful features of bayesian stats (and I'd have loved every word)

0183 18

Riccardo Fusaroli @fusaroli.eurosky.social · 27/06/2026

I guess I've officially entered my contrarian age. Writing a paper deconstructing the notion of speech marker of schizophrenia and having a tad too much fun

041 4

Riccardo Fusaroli @fusaroli.eurosky.social · 23/06/2026

my prize for finishing an article section today was to attempt a representation of la plis, but I got out of tokens before I could get the machine to represent frieren in the world as the fold. :-/

110 1

Riccardo Fusaroli @fusaroli.eurosky.social · 17/06/2026

as far as hjelmslev could (not really that much, it must have been terrible to be one of those students!), it's the equivalent of the duolingo for structuralism (ht @gabrieleferri.bsky.social )

010 1

Riccardo Fusaroli @fusaroli.eurosky.social · 17/06/2026

my new theory: frieren is on a continuum w Lars. Lars uses a doll to re-structure his socially contextualized emotions, but eventually internalizes them (and celebrates a funeral for the doll before moving on with a more healthy emotional life).

110 1

Riccardo Fusaroli @fusaroli.eurosky.social · 11/06/2026

Even in presence of a clear mechanistic link, accuracy drops to clinical uselessness, even w improved algorithms. Recording contexts and equipments change, patients and controls are less clear and more complex cases, more variation in task performance... 5/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 11/06/2026

A. Tsanas showed quite persuasively that having a good mechanistic model of how motor control degeneration impact voice production can lead to v high accuracy in assessing parkinson severity from prolonged phonations (say aaaaaaaaaaaaa) 2/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 09/06/2026

this paper might actually end up being fun to write

010 1

Riccardo Fusaroli @fusaroli.eurosky.social · 09/06/2026

drafting up a new paper bashing some of my old papers, and mentioning cool research along the way. (a new paper on acoustic markers of autism, empirically and conceptually arguing the construct itself is problematic) @kbottemabeutel.bsky.social @irisnomikou.bsky.social

2133 13

Riccardo Fusaroli @fusaroli.eurosky.social · 01/06/2026

it's pretty "fun" to see the pipeline catching likely confabulations (as it was fun to see peer2peer downloads progressing in the 90s?)

020 2

Riccardo Fusaroli @fusaroli.eurosky.social · 31/05/2026

My brother cultivates amazing chickpeas (here the black variant), which I adore, but the kids despise. So I regularly make an Italian version of felafel for my pocket lunch. Delicious

1110 11

Riccardo Fusaroli @fusaroli.eurosky.social · 25/05/2026

Manual annotations across 6638 clips: ~zero correlation between key-child and other-children vocalizations (r = -0.01). VTC says r = 0.39 (medium). LENA says r = 0.07 (weak). Same audio, three very different stories. And the differences track each algorithm's confusion patterns. 5/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 25/05/2026

Here's the key insight: imagine a classifier sometimes confuses female adults with children, then more female adult speech → more "child" speech detected, even when the child is silent. Misclassification doesn't just add noise — it creates correlations that aren't really there. 4/

200 2

Riccardo Fusaroli @fusaroli.eurosky.social · 25/05/2026

Scale up behavioral measurement w ML& you get extra noise & bias. "Classification errors distort findings in automated speech processing," by @lucasgautheron.bsky.social, Kidd, Malko, @marvinlavechin.bsky.social & @acristia.bsky.social shows how to deal w that: pubmed.ncbi.nlm.nih.gov/42151656/ 1/

2163 16

Riccardo Fusaroli @fusaroli.eurosky.social · 20/05/2026

love how this sketch (thought up while swimming) now turned into a paper (and proper graphics). Pre-print soon to appear!

070 7

Riccardo Fusaroli @fusaroli.eurosky.social · 02/05/2026

To show how to concretely use these fingerprints, we simulate each family & use Approximate Bayesian Computation with random forests to recover the generating mechanism from data. The fingerprints are recoverable. Each model fits its own data well & systematically misfits data from the other 2. 10/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 02/05/2026

Yet, each conceptually distinct family leaves distinct, recoverable fingerprints: bimodal vs. tight unimodal vs. exponential inter-call intervals; persistent jitter vs. smooth convergence vs. structureless scatter in phase. Same alternation pattern, very different generative process. 9/

100 1

Riccardo Fusaroli @fusaroli.eurosky.social · 02/05/2026

And the families show striking taxonomic associations: event-triggered dominates in Orthoptera & Anurans, tempo-tracking shows up in Primates, Anurans & Aves, state-dependent in Cetaceans & some primates/birds. But crucially studies almost never test these models against each other. 7/

110 1

Riccardo Fusaroli @fusaroli.eurosky.social · 02/05/2026

How do non-human animal scholars formalize and study the temporal coordination of communicative signals?We did a PRISMA-guided systematic review. From >14k records, 36 studies survived inclusion: formal models of inter-individual timing in the acoustic modality, across 7 taxonomic groups 5/

120 2

Riccardo Fusaroli @fusaroli.eurosky.social · 02/05/2026

"A Systematic Review of Formal Models of Communicative Coordination Across Species" — led by @chrismmcox.bsky.social w @andrearavignani.bsky.social , S.Sørensen, @annapzamm.bsky.social , @frantsjensen.bsky.social , G.Bryant, @acristia.bsky.social, and me - osf.io/preprints/ps... A thread 1/

a panel of figures representing the three model families.

1398 39