Harrison Ritz

@hritz

cybernetic cognitive control computational cognitive neuroscience Connected Minds Fellow @ Queens + RRI he/him 🇨🇦 harrisonritz.github.io

4.497 Folgt4.201 Follower1.174 Posts

YouTube Bluesky

Ist das dein Profil?Übernimm es — dann gehört dir hritz.dirty.social.Profil übernehmen

Harrison Ritz@hritz· 17d

On Thursday, I’ll present a spotlight poster reanalyzing the CCN 2023 paper reviews using Bayesian item response theory 👀

Slide from my poster describing the item response theory model: review ~ N(discrim*quality + bias, discrim^2controversy^2 + noise^2)

Harrison Ritz@hritz· 17d

On Tuesday, @xiaoyiliu.bsky.social has a contributed talk & spotlight poster exploring how abstract task sequences are represented in RNNs and human brains, based on our new OPM-MEG dataset 🤯 Definitely check it out!

Talk title: Orthogonal Task Representations Enable Flexible Task Switching in Human OPM-MEG and Neural Networks

Harrison Ritz@hritz· 27d

so so good

Harrison Ritz@hritz· Jun 19

Very cool speech MEG setup by @joannatoronto.bsky.social

Harrison Ritz@hritz· Jun 4

Same temporal filters, though they are doing extra spatial filtering for the OPM. They also used the same head position indicators for both systems, which will help match their coregistration.

Harrison Ritz@hritz· Jun 4

Looks excellent. Got a chance to see a Fieldline OPM system last week and was super impressed.

Harrison Ritz@hritz· Jun 3

If you’re at #CSBBCS today, scurry back after lunch to catch our symposium on neural network models of brain & behaviour

Harrison Ritz@hritz· May 17

Welcome back to Toronto: winter until May and an absolute unit of a raccoon in your backyard 🥲

Harrison Ritz@hritz· Apr 6

First time in years we’ve gotten to join my family’s Easter pierogi-fest! Love being back close to home :)

265 homemade pierogis on a kitchen counter

Harrison Ritz@hritz· Mar 25

Is that Biscotte? It was unbelievably good

Harrison Ritz@hritz· Mar 23

👼 just wanna be clear how much I dig Petrides & Pandya

Harrison Ritz@hritz· Feb 9

halfway between @alexanderhuth.bsky.social & @gershbrain.bsky.social 🤗 (search is in the top left)

Harrison Ritz@hritz· Feb 3

In conclusion, we find several pitfalls that a decision scientist can encounter when modelling value-based choice, with model misspecification easily producing illusions of control. Some of these pitfalls have been explored in excellent work from perceptual DM (e.g., www.nature.com/articles/nco...)

major pitfalls include participant strategies introducing task confounds, neglecting variants within decision architecture, or failures to consider alterantive decision architectures.

Harrison Ritz@hritz· Feb 3

In accumulator models like the LCA, overall value effects can emerge through a combination of competition and the choice rule. An LCA fit the behavior well. When we fit DDMs to LCA simulations, we found that they reproduced the putative patterns of threshold adjustment.

LCA models do a good job capturing choices, and reproduce the spurious threshold-control effects

Harrison Ritz@hritz· Feb 3

While difficulty no longer had consistent effects on threshold, we now saw that the overall set value did. These were confounded due to how participants weighed different outcomes. Thankfully, a recent project by Romy Froemer helped us understand this pattern. www.nature.com/articles/s41...

If anything, difficult effects on threshold reversed.
But now there were larger effects of overall value on threshold.

Harrison Ritz@hritz· Feb 3

If people used a collapsing bound to meet the deadline, then easier trials would hit the bound at higher point, and harder trials would hit the bound at a lower point. This could mimic the effects of difficulty-dependent control. Indeed, models with collapsing bounds fit behaviour much better.

collapsing bound DDMs fit the behavior much better

Harrison Ritz@hritz· Feb 3

Re-analyzing their data, we replicated their threshold effect, however, the published model didn't fit the data very well. Thankfully, had already published our predictions for what was really going on. osf.io/preprints/ps...

The original model has lower threshold on harder trials, but this model doesn't capture the data well.

Harrison Ritz@hritz· Feb 3

They operationalized vigor through the threshold in a drift diffusion model, predicting that participants would lower their threshold on hard trials in order to respond within a short deadline. This prediction was critical for their EVC model to make predictions for difficulty signals in dACC.

task: participation chose between bundles of outcomes within a deadline

DDM: their model proposed that vigor reduces the threshold on hard trials

Harrison Ritz@hritz· Feb 3

To test the expected value of control (EVC) theory, Vassena et al developed a model in which control 'invigorates' correct responses. www.nature.com/articles/s41... This is a novel prediction from EVC, but it doesn't really make sense: why use control if you already know the right answer?

the log odds of a response depend on value difference plus effort.

effort maximizes the value of the response, minus an effort cost

Harrison Ritz@hritz· Feb 1

This is a craziest scene in Police Story

Scene from Jackie chan’s police story. He is on a hill looking over Hong Kong. Since it is 1985, Hong Kong is much smaller than it is now

Harrison Ritz@hritz· Jan 20

Think it still falls under cognitive science. Original definition were pretty explicit about this relationship. Why not “cognitive science of AI’? Maybe that isn’t catchy enough for granting… rodsmith.nz/wp-content/u...

But the hardware of computers is very far from being the only matter relevant to their
functioning. In order to understand how a computing system works one must enquire into the
logic of the system software and the semantics of the programming languages in which the
system can be addressed. The corresponding questions about human beings are those
asked by the science of psychology - though admittedly, psychological theories seldom
attain a degree of sophistication worthy of their subject matter. An outstanding exception to
this stricture is the science of linguistics, and perhaps it is no coincidence that the most
impressive achievement of artificial intelligence to date is a working model of the
comprehension of natural language.
I would go further and hazard the prediction that for some time to come the most valuable
work in artificial intelligence will be that which attempts to express, in the form of computer
programs, abstract theories of our various cognitive faculties, rather than mathematical
models of the brain itself - this in spite of some excellent recent work on the possible role of
the neocortex as a classifying device. This view is based not only on the obvious vitality of
current artificial intelligence work on language and vision, but also on an evident
dissatisfaction among psychologists with the naive stimulus-response theory of behaviour as
it has been applied to human beings.
It is now plain that a central problem in cognitive psychology is to understand how our
knowledge is represented and deployed, and the computer program is the only medium
which at present offers us the possibility of formulating adequately sophisticated theories of
cognition. The elimination of inadequate theories is no longer the main problem; the defects
of a programmed theory become immediately apparent as soon as it is run on a computer.
In short whatever the technological prospects of artificial intelligence, its principal scientific
value, in my view, is that it sets …

Harrison Ritz@hritz· Jan 1

Spent our New Years apartment hunting Looking forward to 2026!

Harrison Ritz@hritz· Dec 26

Roshan Cools has great work in this area too :) repository.ubn.ru.nl/bitstream/ha...

Harrison Ritz@hritz· Dec 26

it’s a different mechanism (like engine horsepower vs cost of gas) Follows a broader shift in how we think about executive functioning in general, and its role in mental health (including ADHD IIUC) Heres an analogous theory for depression escholarship.org/content/qt3z...

Harrison Ritz@hritz· Dec 22

ok now we’re talking

Harrison Ritz@hritz· Dec 20

Harrison Ritz@hritz· Dec 8

we’re working with Québécois gummies, hbu?

Harrison Ritz@hritz· Dec 6

The issue is how to know when you’re decoding the functional representation (for x, eg decision-making) vs a confound? A gold standard is causal perturbation. Eg Salzman, Britten, & Newsome showing that stim enhances responses to the neuron’s preferred side.

Graph showing that monkeys are more likely to generate a response when side-specific neurons are stimulated

Harrison Ritz@hritz· Nov 25

They summarize it nicely here. I want to read this more closely, and I hope that I haven't misrepresented them. But there are plenty of other issues with PCA-based analyses (e.g., its a decomposition, not a latent variable model; sensitive to temporal jitter; doesn't linearize manifold)

$Our analyses show that during a decision-making task, coding subspaces do not autonomously generate the observed dynamics. Although these dimensions account for a sizable fraction of the population variance, they depend on interactions with other, often weakly selective and low-variance residual dimensions. These residual dimensions lie outside task-coding subspaces and are typically dismissed as noise or irrelevant variability in dimensionality-reduction analyses. However, our results show that they actively drive the dynamics, and perturbing them can flip the choice outcome. The mechanism for this phenomenon is non-normal amplification of neural trajectories that occupy a relatively high number of latent dimensions (∼10). Weak selectivity along the residual dimensions is dynamically amplified through non-normal interactions in the underlying network. As a result, even small perturbations along residual dimensions, which are dominated by neurons with low firing rates and weak task selectivity, can transiently grow and redirect trajectories within the coding subspace. This mechanism explains how seemingly untuned or weakly selective populations can exert strong$