Honey, we fixed Signal Detection Theory (SDT)! In this preprint, Constantin Meyer-Grant, David Kellen, Sam Harding, and I critically evaluate the (unequal-variance) Gaussian SDT model in recognition memory and pursue the Gumbel-min model as a principled alternative: doi.org/10.31234/osf...
🧵

What if forgetting was not the result of time erasing memories but interference from new, similar experiences?

It was what Dr Raina Zhang from the Complex Human Data Hub calls a “mind-blowing” theory that set her on the path to study the mechanisms behind memory.

Learn more: go.unimelb.edu.au/ut7p

Huge thanks to my PhD supervisors @adamosth.bsky.social and @danfeuerriegel.bsky.social for their incredible support throughout this project and my PhD! Special thanks to @nunezanalyzed.bsky.social for showing this method and generously sharing code—this work wouldn’t have been possible without it!

Our findings address the mechanistic account of the LPC overlooked by previous research, and corroborate with the mnemonic accumulator hypothesis (Wager et al., 2005), suggesting the parietal activity during memory retrieval reflects an integration of mnemonic evidence via stochastic accumulation.

As validation, LPC amplitude did not relate to trial-by-trial variation in non-decision time, and the early visual P1 component was unrelated to drift rate. These findings support reinterpreting the LPC as a neural signature of mnemonic strength in evidence accumulation.

By estimating how much LPC variance explained by the model’s cognitive parameters, we showed pre-response LPC amplitude corresponds to trial-by-trial variation in drift rate, signifying memory strength. This link was stronger for previously seen objects and grew stronger as the response approached.

Here, we formally replicated these LPC findings in a new dataset and tested the role of LPC in mnemonic accumulation by jointly modelling behaviours and LPC amplitudes. This was done under a Diffusion Decision Model framework using BayesFlow—a neural network tool for likelihood-free inference.

Crucially, Sun et al. (2024) redefined the LPC measurement, revealing features akin to an evidence accumulation signal (Centro-parietal Positivity). The LPC ramps up and peaks before the recognition response, and early evidence suggest its amplitude varies with memory strength and reaction times.

The Late Positive Component (LPC) is a well-known EEG correlate in recognition memory tasks. Its amplitude reliably tracks recognition performance, and this component is often linked to a high-threshold (all or none) recollection during memory retrieval.

🧠 New preprint alert!

In this study, using a joint modelling method with the Diffusion Decision Model, we offer a mechanistic reinterpretation of the Late Positive event-related potential Component (LPC) as a neural signature of mnemonic strength during evidence accumulation in recognition memory.

We recently updated our preprint that explains how to deal with unidentifiability constraints when measuring participants' decision-making cognition as well as introducing new methods to measure participants' decision-making cognition from brain+behavioral data
osf.io/preprints/ps...

Huge thanks to my supervisors @adamosth.bsky.social and @danfeuerriegel.bsky.social for their continuous support on this project!

Therefore, we suggest that while the variability assumption is meaningful for theories of decision-making, it should not be the only mechanism for slow error predictions in DDM for its estimates to be meaningfully interpreted

We tried to account for this random variability by supplying trial-level endogenous and exogenous drift rate regressors from a large recognition memory dataset with EEG recordings. While the random variability could be accounted for with simulation, this was not observed with experimental data.

This assumption helped the model to account for slow errors and asymptotic accuracy. However, it was criticised for being difficult to estimate and ad-hoc.

DDM is perhaps the most successful evidence accumulation model to account for accuracy and reaction time distribution in decision-making tasks. Ratcliff (1978) proposed that drift rate should vary across trials due to varying levels of item difficulty, which is sampled from a normal distribution.

🚨New Pre-print is out!

What causes the drift rate to vary across trials? How much does the drift rate variability estimate in the Diffusion Decision Model reflect the true variability? Here, we critically examined this by including trial-level regressors of drift rate.

osf.io/preprints/ps...