iiinova — where your digital self keeps evolving

See the full 6 pages for perceptual and meta-perceptual learning effects, prospective confidence, and one very honest closing paragraph:

Finally, while Exp. 2 confirmed that confidence in
absence is affected by stimulus size in a way that parallels
the effect of size on prospective confidence, it also produced
a series of disappointing null findings, with no significant
effect of size on confidence in presence and no correlations
between size effects on post-decisional confidence and on
prospective confidence or perceptual sensitivity.
Disappointed by these null findings, we reminded ourselves
that Exp. 2 had a smaller sample size and only 8 trials per
block (half the number of trials in Exp. 1). As such, we are
not confident that these correlations are truly absent — we
cannot be certain that we would have detected them if they
were present.

140

Despite not being fitted to confidence ratings, model parameters correlated with confidence across subjects. Specifically, a "true visibility" parameter, derived from decisions and RTs, predicted confidence in presence, and a "believed visibility" parameter predicted confidence in absence.

We also fitted our ideal-observer model from psycnet.apa.org/fulltext/202..., in which agents observe noisy sensor activations and decide between PRESENT, ABSENT, or WAIT, as a function of their impatience and, crucially, their beliefs about the expected visibility of stimuli.

A) A schematic description of the ideal observer model. B) Comparison of human and simulated data for Exp. 1, for response time, error rate, and confidence. Lines and points represent human data; semi-transparent rectangles represent data simulated from model parameters, which were fitted to reaction time and accuracy (but not confidence) data of individual participants. Rectangles are centred at their mean value, and their height is twice the standard error.

120

The beautiful thing is that confidence in presence tracked true visibility (higher for smaller stimuli), but confidence in absence tracked believed visibility (higher for bigger stimuli). Only the effect of size on confidence in presence correlated with its effect on objective sensitivity.

A) visual sensitivity (measured as d’) as a function of stimulus size. B) mean confidence in correct responses as a function of target presence and size. Error bars represent the standard error; ***: p<.001. C) correlations between the effect of size on perceptual sensitivity and on confidence, separately for target-present and target-absent decisions.

130

Maya and I used a meta-perceptual illusion to obtain a clear behavioural dissociation. In our task, targets were more visible when the display was small, but participants systematically believed that bigger stimuli made targets easier to see.

Experiment 1. A) Stimuli. Left: example frames from target-present (blue) and target-absent (red) trials. Stimuli refreshed at 15 frames per second until a response was recorded. Right: on different trials, stimuli appeared big or small. C) trial structure. Participants performed two 16-trial detection blocks in which the target was the letter S and two blocks in which the target was the letter A. Stimulus presentation was preceded by a frame, and followed by an analog confidence rating scale. The order of the two letters was randomised between participants.

Confidence in absence as confidence in counterfactual visibility: a CogSci proceedings paper with star MSc student Maya Schipper, is now out on PsyArXiv: osf.io/preprints/ps... 🧵👇

35618

The paper also features fun things like reverse correlation, backwards induction, TAYLOR SWIFT, a new non-parametric sign-consistency test @itayyaron.bsky.social, and some evidence that the Bayesian p(correct) is not a bad approximation of subjective confidence in this task.

Some stuff from the paper: reverse correlation, sign consistency, model predictions for subjective confidence, equations, and taylor swift

2120

Third, we demonstrate this principle with partial occlusion. The more we occluded the display, people missed more targets. But occlusion also made *some* people more likely to make “present” responses on target-absent trials: a reflection of their belief that they may have been missing the target.

A figure detailing the experimental paradigm, in which participants made decisions about the presence or absence of letters under different levels of stimulus occlusion.

161

Second, we show that a model parameter that controls the rate of perceptual evidence affects decisions when a target is present, but it is a parameter that controls *meta-perceptual beliefs about the rate of perceptual evidence* that controls decisions when a target is absent.

A table of correlations between model parameters and occlusion effects on accuracy and reaction times, showing that correlations are significant between the parameter that controls the true visibility of stimuli and behaviour on target-present trials, and between the parameter that controls the believed visibility of stimuli and behaviour on target-absent trials.

150

First, we propose a new model of perceptual detection as a series of momentary decisions between 3 actions: decide present / absent, or accumulate more evidence. The perceptual evidence that goes into this process is noisy, but the decisions are optimal for maximising discounted long-term reward.

An ideal observer model of visual detection

160

Matan Mazor @matanmazor.bsky.social · 15/03/2025

Our "I would have seen it if it were there" paper — a collaboration with @ranimo.bsky.social and @clarepress.bsky.social — is now out in Psych. Review. There’s a lot in this paper, but here are what I see as the 3 main takeaways: psycnet.apa.org/fulltext/202...

The title of the paper, "Beliefs about Perception Shape Perceptual Inference: An Ideal Observer Model of Detection", with the author names, and Fig. 1: a schematic of an inverse optics account of vision.

811441

Kahneman took his science seriously. www.themarker.com/wallstreet/2...

130

Matan Mazor @matanmazor.bsky.social · 14/03/2025

Revising an ethics application, and making sure not to make the participants that I am experimenting upon feel like they are being 'experimented upon'

2182

Matan Mazor @matanmazor.bsky.social · 20/11/2024

From June Huh's Fields Medal interview

While browsing through arXiv postings, rambling between shelves of libraries, or sitting in the dining room of an institute, I met heroes and heroines. The last decade of my life was like living a chapter from ancient mythology. There were hundreds of characters, each one of them with a unique set of abilities. It was, and still is, a privilege to have connected with their minds. On clear days, I can see that I am a small and simple part of a big and complex ancient structure,
like a mushroom sprouted from a giant underground fungal network. It feels good.

031

Matan Mazor @matanmazor.bsky.social · 03/11/2024

George Ainslie on clarity in scientific writing

020

10/11 subject-level model-fitting revealed a beautiful dissociation: while occlusion effects on accuracy and reaction times in target-present trials were a function of visibility, the same effects in target-absent trials were driven by beliefs about visibility.

@ItayYaron @Liad_Mudrik @n_faivre 10/11 subject-level model-fitting revealed a beautiful dissociation: while occlusion effects on accuracy and reaction times in target-present trials were a function of visibility, the same effects in target-absent trials were driven by beliefs about...

100

9/11 (to show that this was true variation and not just noise we used the sign-consistency R package that superstar Itay Yaron developed, together with @liadmudrik.bsky.social , @nfaivre.bsky.social and myself: github.com/mufcItay/sig...). Read more here: osf.io/preprints/psyarxiv/ebg8w

9/11 (to show that this was true variation and not just noise we used the sign-consistency R package that superstar @ItayYaron developed, together with @Liad_Mudrik, @n_faivre and myself: github.com/mufcItay/signcon). Read more here: osf.io/preprints/psyarxiv/ebg8w

100

7/11 so if decisions in target-absent trials are sensitive to occlusion, this is due to something other than visibility itself. According to our model, this something is beliefs about visibility.

7/11 so if decisions in target-absent trials are sensitive to occlusion, this is due to something other than visibility itself. According to our model, this something is beliefs about visibility.

100

6/11 To isolate the role of beliefs about visibility, we occluded parts of the display in a detection task. Our idea was the following: occlusion affects the visibility of present stimuli, but it can’t affect the visibility of absent stimuli.

100

5/11 We provide a Bayesian POMDP model that formalizes this idea, and show that our model captures key behavioural signatures of decisions about absence, assuming perfectly rational, unbiased behaviour.

100

2/11 For such inference to take place, we need access to the “perceptual likelihood function”: a probabilistic mapping from states of the world to sensations. This way, our meta-perceptual model of our own sensory system can affect perception itself.

How do we infer perceptual absences? Do beliefs about perception play a role in such inferences? And how does it all relate to cellist Natalia Gutman’s career decisions? Find answers in our new preprint with Rani Moran and @clarepress.bsky.social. osf.io/preprints/ps... A 🧵

172

Still, pretend games were more predictable and less sensitive to incoming evidence than normal games, and, in the case of Battleship, they were marked by over-acted decision-time profiles – all expected if pretenders were simulating a simplified, “cartoon” model of the decision-making process. (6/7)

100

Pretenders also convinced peers of their ignorance, accurately simulating not only the number of guesses, but also their content and timing, including subtle features such as the quadratic effect of uncertainty on decision times(!) (5/7)

100

Turns out that feigning ignorance is something that people can do impressively well. Check out this correlation between pretend and non-pretend #misses across Hangman target words. Pretenders knew that it would be hard to guess the body part “tooth”, but easy to guess the number “eleven”. (4/7)

100

We devised two new games in which people had to pretend not to know something. In “pretend Battleship”, players were shown where ships were hidden in a grid, but tried to play as if they weren’t. In “pretend Hangman” they pretended not to know the hidden word. (3/7)

120

000

Still, pretend games were more predictable and less sensitive to evidence than normal games, and they were sometimes marked by over-acted decision-time profiles – all expected if pretenders were simulating a simplified, “cartoon” model of the decision-making process. (6/7)

100

100

100

Turns out that feigning ignorance is something that people can do impressively well. Check out this correlation between pretend and non-pretend # misses across Hangman target words. Pretenders knew that it would be hard to guess the body part “tooth”, but easy to guess the number “eleven”. (4/7)

100

Turns out that feigning ignorance is something that people can do impressively well. See this correlation between pretend and non-pretend # misses across Hangman words. Pretenders knew that the body part “tooth” would be hard to guess, but the number “eleven” easier. (4/7)

100

100