Sign in

Arthur Prat

@arthurpr4t.bsky.social
181 followers 90 following 37 posts

Computational cognitive neuroscience. Perception, representation, inference and decision-making. Postdoc at Harvard with Sam Gershman. arthurprat.com

PostsRepliesMedia
Arthur Prat @arthurpr4t.bsky.social · 27/09/2026
Finally, I'd be curious to know whether the differences you get might come from task-specific representations, as I argue. In any event, thank you! Great study.
100
Arthur Prat @arthurpr4t.bsky.social · 27/09/2026
Maybe this could reduce the gap you have between the two measures. Also you could fit one model to both datasets, and see whether having the same or different parameters across tasks improves the BIC or some other measure.
100
Arthur Prat @arthurpr4t.bsky.social · 27/09/2026
Perhaps it makes a small difference; but more importantly when computing D(x) I would not include this sigma_0, on the argument that it is some other kind of noise that should not impact the discrimination behavior (but maybe you already do this?).
100
Arthur Prat @arthurpr4t.bsky.social · 27/09/2026
A few quick comments: for the noise (Eq 3) it seems more natural to me to sum variances (instead of stddevs). That way sigma_0^2 could be interpreted as some other noise that just adds to the representation noise (I would say motor noise but with oral reports that's less obvious).
200
Arthur Prat @arthurpr4t.bsky.social · 27/09/2026
Thank you! Yes I think this is very interesting. (Are there any other studies where both estimates and discrimination are elicited??)
200
Reposted by Arthur Prat
gillesdehollander.bsky.social @gillesdehollander.bsky.social · 21/09/2026
New paper in Nature Communications 🧠 Rapid Changes in Risk Attitudes Originate from Bayesian Inference on Parietal Magnitude Representations. Why the same person can flip between a safe bet and a gamble, and what parietal cortex has to do with it. 🧵 www.nature.com/articles/s41...
nature.com
Rapid changes in risk attitudes originate from Bayesian inference on parietal magnitude representations - Nature Communications
Risk attitudes can shift across contexts and even between identical choices. Here, the authors show that as parietal payoff representations grow noisier, the brain leans more on prior beliefs, biasing...
1217
Arthur Prat @arthurpr4t.bsky.social · 17/09/2026
Nice. I wonder whether an LLM trained on pre-1900 data could come up with quantum physics, as Planck did. That would tell us something about LLMs' abilities.
010
Arthur Prat @arthurpr4t.bsky.social · 17/09/2026
Thank you! Super interesting indeed. What I find conceptually very very nice is the notion of the trajectory of the precision over episodes, and that you might want to start by "paying" more for high precision, before lowering. I wonder if one can get analytical results, maybe in simpler settings.
010
Arthur Prat @arthurpr4t.bsky.social · 12/09/2026
Thank you!! It is really great that you find task-specific representations. (I wonder why the similarity task should have this specific representation. Presumably the specific objective of this task, combined with a prior over orientations that favors cardinal directions). I'll have a look at Smith.
110
Arthur Prat @arthurpr4t.bsky.social · 09/09/2026
as a function of task demands. So representations depend on what they are used for. To understand cognition and decision-making, we must first understand representations! (because that's what cognition and decision-making operate on).
180
Arthur Prat @arthurpr4t.bsky.social · 09/09/2026
This is not the first efficient-coding proposal out there, but the originality of this one — and why it explains so many results at once — is that it includes an 'endogenous' component, in which the degree of noise is itself a decision variable, i.e., the brain can choose how imprecise to be,
1100
Arthur Prat @arthurpr4t.bsky.social · 09/09/2026
the sublinear scaling laws of the imprecision with prior width (which I had reported in earlier work); Shepard's universal law of generalization; and a quantile-invariance property of the estimation bias, which has not been reported before.
170
Arthur Prat @arthurpr4t.bsky.social · 09/09/2026
We derived six laws of psychophysics from a single efficient-coding equation. The laws include Weber's law; scaling laws in visual working memory (which had been noted before, but with unclear theoretical justification); Wei & Stocker's law of human perception; arthurprat.com/pdfs/Prat-Ca...
arthurprat.com
49628
Reposted by Arthur Prat
Vinny Costa @vincentcostaphd.bsky.social · 07/09/2026
More stories like this one, please and thank you www.nytimes.com/2026/09/06/s...
nytimes.com
The 92-Year-Old Mathematician and the Teenage Apprentice
Joan Birman thought her major discoveries were behind her. Then came an email from a young neighbor — a girl who knew little but wanted to learn.
0189
Arthur Prat @arthurpr4t.bsky.social · 21/08/2026
We broke Weber's law. By manipulating the prior (ie the frequencies of small and large magnitudes), we changed how people's accuracy depended on magnitude. When large magnitudes were more frequent, subjects became more precise about them. This points to efficient coding, dynamically implemented.
15515
Reposted by Arthur Prat
Blair Shevlin @bshev.bsky.social · 20/08/2026
Now published in eLife: elifesciences.org/articles/103... We find strong evidence that activity in Pre-SMA and vmPFC scales with gaze-weighted accumulated evidence, suggesting attention directly modulates value signals in canonical decision-making circuits! 🧵 on the methods + findings:
elifesciences.org
Overt visual attention modulates decision-related signals in the frontal cortex
Brain activity in the pre-supplementary motor area and dorsolateral prefrontal cortex represents gaze-weighted accumulated evidence signals in value-based decision-making.
1226
Arthur Prat @arthurpr4t.bsky.social · 21/08/2026
Indeed, and we touch upon that in the paper. In some cases the JND is a non-monotonic function of magnitude. But it's still conceived as a fct of the magnitude. Here we break this relationship between JND and magnitude by manipulating the prior. Presumably the real relation is between JND and prior
120
Arthur Prat @arthurpr4t.bsky.social · 21/08/2026
With the large-dominant prior, large magnitudes are perceived more precisely, so the second effect (bias toward first stimulus, with small magnitudes) becomes more salient. The paper focuses on these changes where we manipulate the prior, rather than on order effects.
010
Arthur Prat @arthurpr4t.bsky.social · 21/08/2026
Conversely, for small magnitudes, choices are biased toward the first one. But the sizes of these effects are different in the two conditions. With the more natural small-dominant prior, the first (large-magnitudes) effect is more salient, probably because there is more noise in large magnitudes.
110
Arthur Prat @arthurpr4t.bsky.social · 21/08/2026
Thanks! Yeah there is an order effect, essentially consistent with the idea that the first one is perceived with more noise and thus estimated as closer to the mean. So for large magnitudes, the second one is chosen more often than it should, presumably because the first one is under-estimated.
110
Reposted by Arthur Prat
samuel mehr @mehr.nz · 16/07/2026
calling all researchers who collect data from humans 🧠🧪 we are running a very brief survey (<1 min) about jsPsych, the software ecosystem for browser-based data collection. if you use jsPsych, or if you collect data with any other software, please fill it out at tinyurl.com/jspsych-census!
I am writing on behalf of the leadership of jsPsych, the software ecosystem for browser-based data collection used in experiments and surveys from many labs worldwide.

It has been difficult for us to figure out how widespread jsPsych usage is, because the software is freely available online and we do not collect any information about downloads, users, experiments posted online, etc. While the software has been cited in thousands of articles, jsPsych usage is likely more widespread than those citations imply, because not every project using jsPsych leads to a citation, and because there are also often substantial delays between actual jsPsych usage and a publication. As such, we are conducting a very brief census of jsPsych users.

If you collect data from humans using jsPsych or any other type of software, we would appreciate your filling out a very brief survey (<1 min) at tinyurl.com/jspsych-census. This will help us to understand how many people, labs, and institutions are current or former users of jsPsych, information that will help support the long-term sustainability of the jsPsych ecosystem.

We would also appreciate it if you could circulate this message to your lab, your department listserv, and your collaborators, to help ensure that it reaches as many members of the research community as possible. The survey is being distributed by Josh de Leeuw, Melissa Kline Struhl, and myself. Please contact me (sam@auckland.ac.nz) off-list if you have any questions about it.
13855
Arthur Prat @arthurpr4t.bsky.social · 26/06/2026
Oh that's nice to hear! Of course I'm happy to talk about, if helpful
020
Arthur Prat @arthurpr4t.bsky.social · 25/06/2026
To give a clear view of what's going on, throughout I did my best to provide analytical expressions for the important quantities in this somewhat complex system. Hopefully this is helpful! Also, thanks to Reviewer #2.
110
Arthur Prat @arthurpr4t.bsky.social · 25/06/2026
This accounts for both prior attraction and adapter repulsion (including a bunch of subtle effects related to adapter repulsion). We also find behavioral evidence consistent with prior attraction.
110
Arthur Prat @arthurpr4t.bsky.social · 25/06/2026
This is what we look at in this paper finally published a few days ago in Nature Comms. We propose that gain adaptation in recurrent networks optimizes an efficient-coding objective, balancing accuracy and spiking cost. The modulated gains propagate through the network, shifting the tuning curves.
110
Arthur Prat @arthurpr4t.bsky.social · 25/06/2026
To adapt to a new prior, tuning curves should "move" toward more frequent stimuli ('prior attraction'). But in adaptation studies they are shown to move away from the more frequent stimuli ('adapter repulsion'). Oh also how do tuning curves move at all?? www.nature.com/articles/s41...
nature.com
Fast efficient coding and sensory adaptation in gain-adaptive recurrent networks - Nature Communications
How sensory systems rapidly adapt to changing stimulus statistics remains unclear. Here the authors show that gain adaptation in recurrent networks can implement fast efficient coding, unifying prior ...
1235
Arthur Prat @arthurpr4t.bsky.social · 21/06/2026
These days the expected value of the Massachusetts Megabucks lottery is positive, so I expect every economist and decision theorist in Massachusetts will get their ticket. If you don't buy, clearly something must be wrong with your utility function
030
Arthur Prat @arthurpr4t.bsky.social · 20/05/2026
There's no eye in AI
010
Arthur Prat @arthurpr4t.bsky.social · 26/04/2026
by mixing in bond indices and the like. In any event, given that Harvard will presumably outlive all of us, I think its best strategy is a very long term one that accepts big risk for big returns (but obviously the minimal risk for that level of return, which is precisely what indices provide)
130
Arthur Prat @arthurpr4t.bsky.social · 26/04/2026
Now, even investing in indices (which I'm sure the people at Harvard don't do) it would be acceptable to perform less well than the S&P if one wants to avoid huge market swings (e.g. -30%). Perhaps it's reasonable if this finances a sizable fraction of operating costs. But it's easy to de-spice that
110
Arthur Prat @arthurpr4t.bsky.social · 26/04/2026
especially after mgmt fees and trading costs. That's the whole Warren Buffett paradox, who recommends everyone to buy indices, while himself cherry picking stocks (successfully). But given that it is possible, "why not me? I'm sure I can do it" seems to be the usual response.
130
Arthur Prat @arthurpr4t.bsky.social · 26/04/2026
I think the main reason people still turn to active asset management, even though it's been demonstrated again and again and again that indices are just better, is that in fact it *is* possible to beat the market, but it's hard, and active managers cannot collectively succeed,
110
Arthur Prat @arthurpr4t.bsky.social · 21/04/2026
With in mind the view to go after more complex representations, in the future. I thank the reviewers and editors for assessing the findings as important and the evidence as compelling, and more generally @elife.bsky.social for a smooth overall process.
011
Arthur Prat @arthurpr4t.bsky.social · 21/04/2026
Representations are the basis of cognition. But it seems to me that they are not always well understood, or characterized, in cognitive science and other behavioral sciences. This paper is an attempt to do that, and more precisely to get at the principles under the representations of simple scalars.
111
Arthur Prat @arthurpr4t.bsky.social · 21/04/2026
The "Version of Record" of my paper with Mike Woodford is finally there. I am very glad with this paper that points to new laws of psychophysics, which we derived from a carefully considered principle of endogenous efficient coding.
elifesciences.org
Endogenous precision of the number sense
A theory of efficient coding wherein the precision of representations is endogenous, task-dependent, and prior-dependent predicts scaling laws for imprecision and is supported by numerosity perception...
1186
Arthur Prat @arthurpr4t.bsky.social · 15/04/2026
Isaiah Berlin on FDR in the 1930s: "... to construct a regime which should provide for greater economic equality and social justice - ideals which were the best part of the tradition of American life - without altering the basis of freedom and democracy in his country." Where is our Roosevelt!
000
Arthur Prat @arthurpr4t.bsky.social · 02/04/2026
"How much would you pay to not know?" (WTP right to ignorance) "How much would you ask for being forced to know?" (WTA right to ignorance) "How much would you pay to know?" (WTP right to know) "How much would you ask for not being told?" (WTA right to know)
020
Arthur Prat @arthurpr4t.bsky.social · 02/04/2026
Also it is conceivable that some would be ready to pay for knowing, so as to be liberated from the throes of uncertainty. That would add a term in your calculation of the "utility" of knowing. So now you have four numbers, corresponding to the answers to the questions:
120
Arthur Prat @arthurpr4t.bsky.social · 02/04/2026
The willingness to accept is usually larger than the WTP so you might get more positive amounts if you ask "how much do you ask to relinquish your right to ignorance?"...
110
Reposted by Arthur Prat
Pierre-Etienne Fiquet @pfiquet.bsky.social · 26/03/2026
Applications are now open for our Junior Theoretical Neuroscientists Workshop which will take place July 21 - 24, 2026 at the Center for Computational Neuroscience @flatironinstitute.org Learn more and apply by April 15 at www.simonsfoundation.org/event/jrwork...
simonsfoundation.org
Junior Theoretical Neuroscientists Workshop
Junior Theoretical Neuroscientists Workshop on Simons Foundation
1138
Arthur Prat @arthurpr4t.bsky.social · 27/03/2026
Just posted an update of this study, where we show how the receptive fields of number-encoding neural populations shift and rescale with the prior — "distributed range adaptation" — in line with (dynamic) efficient coding, and predictive of behavior. Check it out! www.biorxiv.org/content/10.1...
0276
Arthur Prat @arthurpr4t.bsky.social · 15/03/2026
Great work. I wonder how this relates to www.cell.com/cell-reports... which says something similar. (But with very ≠ setups in terms of modeled tuning curves. Also the metric used is different: you show how the FI scales with D, they show the prob of errors is exponential in the number of neurons.)
cell.com
Random compressed coding with neurons
Blanco Malerba et al. demonstrate that extended, irregular tuning curves, observed in the brain, efficiently compress information in a low-dimensional representation. Optimal smoothness of tuning curv...
041
Reposted by Arthur Prat
Sam Gershman @gershbrain.bsky.social · 17/07/2025
Efficient coding theories often implicitly or explicitly assume slow changes in tuning (e.g., through synaptic plasticity). Arthur Prat-Carrabin has collected psychophysical data showing that it can be fast, and this can be explained by a gain-adaptive RNN: www.biorxiv.org/content/10.1...
biorxiv.org
Fast efficient coding and sensory adaptation in gain-adaptive recurrent networks
As the statistics of sensory environments often change, neural sensory systems must adapt to maintain useful representations. Efficient coding prescribes that neuronal tuning curves should be optimize...
04610