Casper Kerrén

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

There is much more in the paper: sampling-history-dependent competition between memories, degraded-cue retrieval, pattern-completion predictions, and links to cortical-hippocampal memory systems. Data, code and model are on Zenodo: doi.org/10.5281/zeno...

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

Finally, we explore what this might mean for reconstructive memory. We train a generative network model to reconstruct images from the sequentially formed memory representation, explicitly prioritising regions that had been sampled during encoding.

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

What about relations between elements? Holding the sampled information and total exposure to A and B constant, relational retrieval improves as the sampling sequence transitions more frequently between A and B. More opportunities to move between elements strengthen access to their association.

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

Now back to the two people looking at Goya. If memory simply accumulated an unordered set of visual samples, changing their order should not matter. But when we shuffle or reverse identical samples, the final representation changes substantially and exact-image retrieval gets worse.

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

Something more surprising happens to representational geometry. As more samples are integrated, population dimensionality decreases, even though access to the exact image improves. The representation becomes organised along fewer dominant dimensions, without losing item-specific information.

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

That lets us separate two timescales. First, slow learning of how to integrate sequential samples, and second one-shot formation of individual memories. We then ask how sampling shapes both the representation of a single element and relations between elements.

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

Our model never sees the whole image at once. It receives a sequence of local samples, like fixations. Each sample updates an evolving internal state that already reflects what came before, so the same input can have a different effect depending on its history.

Imagine two people standing in the same spot, looking at the same Goya. They inspect exactly the same parts of the painting, but in a different order. Later, they talk about what they saw. Ignoring all other differences: are the memories they formed identical?

Most computational models of episodic memory start once a representation has already been formed and ask how it is stored or retrieved. But how was that representation constructed in the first place from the sequence of partial samples gathered as we look around?

Very happy to share our new paper, now published in @elife.bsky.social, together with @gonzalezgarcia.bsky.social and @lindedomingo.bsky.social elifesciences.org/reviewed-pre...

Simply keeping something in memory for longer did little. But introducing an intervening task selectively increased the semantic advantage in non-decision time.

Very happy to share our new paper, now published in @elife.bsky.social, together with @gonzalezgarcia.bsky.social and @lindedomingo.bsky.social elifesciences.org/reviewed-pre...

When they knew which item would be needed, the semantic advantage in non-decision time was almost completely eliminated. When attention remained distributed across several memories, it was large.

Very happy to share our new paper, now published in @elife.bsky.social, together with @gonzalezgarcia.bsky.social and @lindedomingo.bsky.social elifesciences.org/reviewed-pre...

Semantic judgements consistently had shorter non-decision times than perceptual judgements, across memory loads and delays. Differences in drift rate were much more conditional, emerging particularly under greater mnemonic demands. Semantic advantage appears primarily before evidence accumulation.

Very happy to share our new paper, now published in @elife.bsky.social, together with @gonzalezgarcia.bsky.social and @lindedomingo.bsky.social elifesciences.org/reviewed-pre...

We first revisited our previous working-memory data using hierarchical drift-diffusion modelling. This allowed us to separate two things hidden within reaction time: → processes occurring before evidence accumulation → the accumulation of evidence itself The result was striking.

Very happy to share our new paper, now published in @elife.bsky.social, together with @gonzalezgarcia.bsky.social and @lindedomingo.bsky.social elifesciences.org/reviewed-pre...