Attractor dynamics underlie sequence working memory in data-constrained recurrent neural networks
By fitting low-rank recurrent neural networks to multi-session macaque recordings, this study demonstrates that sequence working memory relies on position-specific coding subspaces where attractor-like dynamics maintain persistent item representations, while activity simultaneously evolves along distinct temporal dimensions.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Imagine trying to hold a short string of numbers in your mind long enough to dial a phone or enter a code. This mental feat, known as working memory, allows us to keep information active for a few seconds after the original stimulus is gone. For decades, scientists have debated how the brain accomplishes this. One leading theory suggested that specific groups of neurons simply keep firing steadily, like a light bulb left on, to hold the information. Another theory proposed that the brain uses a rapid, shifting sequence of activity patterns, where one group of neurons triggers the next in a chain, creating a stable overall picture even though the individual parts are constantly changing. Resolving this debate is crucial because it reveals the fundamental operating system of the brain, determining whether our thoughts are held in place by static anchors or by a flowing river of activity.
A team of researchers at Tübingen University and the Tübingen AI Center has now investigated this question by building a sophisticated computer model of the brain's prefrontal cortex, the area responsible for complex thought and memory. They used data from two macaques performing a memory task where they had to remember a sequence of one to three visual shapes. The monkeys were trained to pull a lever, watch the shapes appear in a specific order, wait for a brief pause, and then touch the screen to reproduce the sequence. The researchers recorded the electrical activity of 2,488 individual neurons across 30 different sessions to see exactly how the brain handled this task.
To make sense of this massive amount of data, the scientists created a "data-constrained" recurrent neural network. This is a type of artificial intelligence model designed to mimic the way real neurons connect and fire. Unlike standard models that are trained only to get the right answer, this model was forced to reproduce the actual, messy patterns of spikes seen in the monkey brains. By fitting the model to the real data, the researchers could observe the hidden dynamics that occur between the neurons, which are impossible to see directly in the raw recordings. The model successfully recreated the statistical properties of the real brain activity, including how often neurons fired and how they influenced one another, confirming that it was a faithful representation of the biological system.
When the researchers analyzed the internal workings of this model, they discovered a specific geometric structure in how the information was stored. The brain activity for each item in the sequence was organized into separate, distinct zones within the neural network. If the monkey remembered a shape in the first position of the sequence, the activity settled into one specific zone; if it was the second or third shape, the activity settled into different, non-overlapping zones. This confirmed earlier observations that the brain keeps track of "what" the item is and "where" it falls in the sequence by using separate spatial areas for each position.
The critical question remained: what happens inside those zones while the monkey waits? Does the activity stay perfectly still, or does it flow? The researchers found that the answer was a combination of both. While the overall pattern of activity remained stable enough to be decoded as a specific memory, the underlying dynamics were not static. Instead, the model revealed that the memory was held within a stable "manifold," a curved surface in the mathematical space of the network. Within this surface, the activity behaved like a ball rolling into a valley and settling at the bottom. This is known as an attractor state. Even though the brain was also processing the passage of time with other, separate signals, the actual memory of the shapes was held by these stable, self-correcting states.
To prove that these states were truly stable and not just a fleeting sequence of events, the researchers performed a virtual experiment. They took their trained model and gave it a tiny, simulated jolt, similar to what might happen if a scientist used light to briefly stimulate a group of neurons in a real monkey. They found that if the jolt pushed the activity slightly off its path, the system naturally pulled itself back to the correct memory state, just as a ball in a valley rolls back to the bottom. However, if they pushed the activity in a specific direction along the stable surface, the model's memory shifted permanently to a new value. This behavior is the hallmark of an attractor system: it resists small errors but allows for controlled changes.
The study suggests that the brain does not rely on a single mechanism for memory. Instead, it uses a hybrid strategy where stable attractor states hold the specific content of the memory, while other parts of the network handle the timing and flow of the sequence. This finding challenges the idea that working memory is purely a transient, shifting chain of events. It indicates that for the brain to hold a sequence of items, it must have stable anchors that keep the information safe from noise, even as time passes. The researchers also noted that their model predicts that if scientists were to apply a brief, targeted perturbation to the brain during the delay period, they could selectively change the monkey's memory of one item in the sequence without affecting the others. This provides a clear, testable prediction for future experiments to verify whether this attractor structure exists in the living brain.
Ultimately, this work bridges the gap between two competing views of memory. It shows that the brain can maintain a stable representation of the past without requiring every single neuron to fire continuously at a constant rate. By organizing memory into specific, stable zones that correct themselves when disturbed, the brain ensures that the information we need remains accessible and accurate, even as the world around us changes. The researchers' approach of building models that are tightly constrained by real biological data offers a powerful new way to understand the hidden rules that govern how we think and remember.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.