iiinova — where your digital self keeps evolving

More broadly, these findings show that statistical learning can remain flexible even when the same item supports different predictions across changing contexts. For more info, please check out this exciting work led by my talented grad student Fleming Peck: www.biorxiv.org/content/10.6...

biorxiv.org

Why do some models succeed? 🗝️The key is HOW THEY REPRESENT CONTEXT. Successful models form more distributed representations: the same input produces different hidden states depending on context. This supports flexible switching and prevents interference.

Only models with moderate weight initialization variance learned both contexts. Too little variance → stuck in the most recent context Too much variance → noisy, unstable representations An intermediate regime enables successful context-dependent learning, and mimicked human performance.

100

We then trained recurrent neural networks (GRUs) on the same task. These models learn to predict the next item in a sequence — allowing us to test how context-dependent structure might emerge from experience alone. No explicit context signal was provided.

Reaction times also revealed online learning. Participants became progressively faster to the predictable second item in each pair over time, consistent with growing anticipation. For context-dependent pairs, early slowing may reflect interference from recent context switches.

100

Participants showed above-chance learning of BOTH associative structures! This happened even when context was never signaled (fully latent). Adding an explicit visual context cue (border color) didn’t help, suggesting learners inferred context directly from sequence statistics.

Participants viewed a continuous stream of unfamiliar objects and made a simple perceptual decision on each one. 🥨The twist: - the same objects predicted different next objects depending on which of two contexts was active No learning instructions. No feedback. Just exposure.

100

Mariam Aly @mariamaly.bsky.social · 03/04/2025

🚨NEW PREPRINT (w/ Fleming Peck & Hongjing Lu) "Spontaneous emergence of context-dependent statistical learning in humans and neural networks" We are constantly predicting what will happen next. Yet the same cue can lead to different predictions depending on context. How might this work...? 🧵1/8

13814

Reposted by Jesse Rissman

My favorite conference is the Memory Disorders Research Society meeting. It's a delightful community: top-notch research & wonderful people who have been so supportive in my career. Want to join? Nominations for membership (including self-nominations) are open until April 9! Form at the top👇🏼

memorydisorders.org

MDRS

MDRS is a professional society dedicated to the study of memory. Members engage in basic and clinical research into how memory works and why it fails.

23515

Reposted by Jesse Rissman

James Antony @jameswardantony.bsky.social · 04/04/2025

The Memory Disorders Research Society (MDRS) has extended its deadline in seeking nominations for new members to next Wednesday (4/9)! Self-nominations welcome. Feel free to DM me if you have questions about the society or the membership process. Application link: docs.google.com/forms/d/e/1F...

docs.google.com

0116

Our results suggest that face perception and memory might be less intertwined than we might think, with implications for prosopagnosia and memory related disorders. Work led by the fantastic Jan Kadlec.

We test this with an additional face inversion experiment. We find greater inversion effects for easier perceptual levels --> more holistic processing at easy levels --> more interaction at easy levels as a result of face-specific interference which disrupts holistic processing.

Instead, we hypothesize that the face-related distractor in the face-specific interference task disrupts holistic face processing, pushing participants to switch to a feature-based strategy instead.

100

This pattern of interaction does not fit with theories of increased cognitive load, or overlapping shared resources. Also, just increasing cognitive load through an orthogonal, non-face related interference condition (math), has negligible effect on performance on the face task.

100

An equally robust finding though, is that under more complex conditions (when an emotional face interference task was inserted into the delay period), we do find interactions. Surprisingly, these go in an unexpected direction - more interaction for easy rather than hard perceptual conditions. Why?

100

Across four large, independent datasets, with over 800 participants in total, we find clear evidence that face perception and face working memory are fundamentally independent, at least when task demands are low.

100

We parametrically modulate difficulty in each domain separately. Shared resources would imply interactions between perceptual and memory difficulty, with increasing interaction as difficulty increases. If they are independent however, we would expect difficulty to be additive.

100

We designed the Face Memory and Perception (FMP) task to systematically test whether perceptual and memory components of face processing rely on shared cognitive resources, or function independently.

100

Behavioural separation of face memory and face perception

Excited to share this new preprint, along with an explainer thread written by my talented collaborator Michal Ramot. Are face perception and face memory fundamentally independent processes, or are they overlapping and impossible to disentangle? 🧵 www.biorxiv.org/content/10.1...

biorxiv.org

A long-standing debate in neuropsychology concerns whether perception and memory function as independent systems or interact to support cognition. To investigate this, we developed the Face Memory and...

152

Jesse Rissman @rissman.bsky.social · 12/12/2024

Still one month left to apply for the tenure-track faculty position in "Measurement Issues in Complex Data Structures" in the UCLA Dept of Psychology. The scope is broad, but includes methodologists working with brain data (fMRI, EEG, cellular neurophysiology/imaging). recruit.apo.ucla.edu/JPF09885

Jesse Rissman @rissman.bsky.social · 17/11/2024

Would love to be added if I’m not already on there. Thanks!

100

Reposted by Jesse Rissman

Paul Scotti @humanscotti.bsky.social · 19/03/2024

🧠👁️ Our MindEye2 preprint is out! We reconstruct seen images from fMRI activity using only 1 hour of training data. This is possible by first pretraining a shared-subject model using other people's data, and then fine-tuning on a held-out subject with only 1 hr of data. arxiv.org/abs/2403.11207

2249

In case you’re interested in trying out these methods, all of @crewalsh.bsky.social’s code is available at osf.io/5q6th/

000

Future studies will be needed to better characterize whether these subtle learning-induced semantic distortions are short-lived or whether they can endure for weeks or months.

100

Despite the general stability of semantic knowledge over the course of one’s lifetime, our results demonstrate that even a brief session of episodic learning can subtly yet systematically re-sculpt semantic space, and this representational change impacts subsequent recall.

Testing, on the other hand, facilitates elaborative connections linking paired words – when these connections don’t already exist (like in semantically related pairs), new connections are created, which subtly change the representation content of both cue and target.

100

Together, our results paint a picture where there are two (interacting) processes occurring during paired associate learning. Pre-existing semantic information enhances memory by reducing interference from potential lures and making cues more predictive of targets.

100

Change in the target, on the other hand, was only predictive of later recall in unrelated pairs that were tested – for those pairs, the more the target representation changed, the more likely the pair was to be recalled.

100

Finally, we looked to see how all this representational change related to behavior. Regardless of how a pair was learned or the semantic relatedness of the pair, more representational change in the cue was associated with a higher probability of recall.

100

Then, we looked at how the representational structure of individual words changed across learning - we showed that semantically related cues were drawn asymmetrically towards their targets, while semantically unrelated cues and targets were drawn together symmetrically.

100

Next, we showed that testing repelled words that were moderately related to the to-be-learned cues. Put another way, words that might have tripped up recall of our to-be-learned pairs were pushed further away in semantic space so they didn’t interfere with recall.

100

Using this measure, we first showed that words in semantically related pairs were drawn closer in representational space than those in semantically unrelated pairs.

100

This gave us a sensitive measure of how the semantic representation of each of our words changed across learning.

The twist here was that we never directly asked them to make judgements on words in our to-be-learned pairs – we didn’t want them to have any top-down biases after learning. That meant that we had to impute these values using the relational judgements of the other words!

Just looking at accuracy doesn’t tell us about how this change happens, so we were inspired by neuroimaging analyses and had our participants do a drag-and-drop task where they judged how related words were to each other before and after learning.

100

At the final test, there was a smaller testing effect for semantically related pairs. Pre-existing semantic information rescued memory for the restudied pairs to make the difference between testing and restudying smaller!

100

We wanted to understand how these processes interact, so we had 80 participants learn semantically related and semantically unrelated pairs of words through either active retrieval practice or passive restudying.

100

BUT how we engage with information can also impact how well we learn it - actively retrieving information from memory tends to strengthen it more than passively restudying it. For instance, using flashcards tends to be a better studying method than re-reading from a textbook

100

A quick summary thread from my recently graduated PhD student Catherine Walsh: Even though episodic and semantic memory are often studied separately, semantic information can provide an important scaffolding to help build new episodic memories - things that are related are often easier to remember.

100

Mariam Aly @mariamaly.bsky.social · 21/11/2023

In case you missed @crewalsh.bsky.social’s thread on the other network, I wanted to share our latest paper here. We used a novel approach to examine the representational underpinnings of the testing effect and the bidirectional interplay of episodic & semantic memory. www.nature.com/articles/s41...

Screenshot of published paper showing title (“Behavioral representational similarity analysis reveals how episodic learning is influenced by and reshapes semantic memory”) and abstract.

1144

Reposted by Jesse Rissman

How does episodic learning affect the structure of semantic memory? Episodic learning pulls associated words together in semantic space. A cue word's semantic representation moves closer to the target word. Clever use of behavior to understand memory representations! By Walsh & @rissman.bsky.social

nature.com

Behavioral representational similarity analysis reveals how episodic learning is influenced by and r...

Pre-existing semantic knowledge provides an organizational structure for episodic memories. Here, the authors show that episodic learning systematically shapes this semantic space depending on how lea...

13725

Reposted by Jesse Rissman

Jaan Aru @jaanaru.bsky.social · 05/10/2023

Hippocampal neurons code individual episodic memories in humans. The neurons don't code for a particular element in the episode (e.g., concept or time). They code for the conjunction of the different elements. #neuroscience #psychology #PsychSciSky #NeuroSkyence www.nature.com/articles/s41...

nature.com

Hippocampal neurons code individual episodic memories in humans - Nature Human Behaviour

Kolibius et al. show that individual neurons in the human hippocampus code for particular episodic memories.

04432

Jesse Rissman @rissman.bsky.social · 03/10/2023

Thanks, I forgot to say that this is for a cognitive psychology course, so the main thing I am trying to convey is how auditory and visual stimuli are presented and how button press responses are collected. I don’t really need to showcase all of the gradient coils.