Firing Without Wiring: How REM-like Sleep Reactivation Rebalances Excitation and Inhibition to Orthogonalize Memories
Using biologically plausible recurrent neural networks, this study demonstrates that REM-like sleep facilitates memory consolidation through a "reactivation-targeted pruning" mechanism, where the reactivation of recent memories selectively weakens their excitatory inputs from older memory populations to reduce interference and orthogonalize representations without requiring inhibitory plasticity.
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 This is an AI-generated explanation of a preprint that has not been peer-reviewed.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
Every night, the human brain cycles through different states of rest, including a phase known as rapid eye movement, or REM, sleep. For decades, scientists have understood that this state is vital for memory, yet the exact mechanical reason why has remained a mystery. We know that during non-REM sleep, the brain replays recent experiences to strengthen them, much like reviewing notes before a test. However, REM sleep looks different; the brain is active, but the patterns of activity do not seem to replay specific events in the same orderly way. Instead, many researchers have long suspected that REM sleep serves a different, perhaps more subtle purpose: it might act as a cleanup crew, separating memories that have become too similar or tangled. If two memories share too many details, they can interfere with one another, making it hard to recall either one accurately. The prevailing theory suggests that REM sleep untangles these connections, ensuring that similar memories remain distinct and do not overwrite each other. Despite this compelling idea, proving exactly how the brain achieves this separation at the level of individual connections has been difficult, as we cannot watch the microscopic changes in a living human brain in real time.
To solve this puzzle, researchers at the University of California San Diego turned to computer simulations that mimic the brain's learning processes. They built a digital brain, a type of artificial network designed to learn tasks one after another, similar to how a student might learn to recognize different pairs of numbers in sequence. In this digital world, learning a new task often caused the system to forget the old ones, a problem known as catastrophic forgetting. The researchers then introduced a simulated sleep phase after each new learning session. This sleep was not a period of rest where the system shut down, but rather a time when the network was reactivated by random noise, mimicking the chaotic firing patterns seen during REM sleep in biological brains. During this simulated sleep, the connections between the digital neurons were allowed to change based on simple, local rules, without any external teacher telling the system what was right or wrong.
The results of these simulations revealed a surprising mechanism for how memories are protected. When the digital brain learned a new task, the neurons responsible for that new task became dominant, effectively hijacking the inputs that previously belonged to older memories. This is what caused the forgetting. However, when the simulated sleep phase began, the neurons representing the most recent task were the ones that fired the most. Because of this dominance, the connections feeding into these active neurons underwent a specific type of change. The inputs coming from the older, silent memories were weakened, while the connections from the recent, active memories were preserved or slightly strengthened. This process effectively pruned the unwanted connections, stripping the new memory of the inputs it had stolen from the old ones. As a result, the older memories were freed up to function on their own again, and the system could recall both the new and old tasks with high accuracy.
This mechanism, which the authors call "reactivation-targeted pruning," works by ensuring that plasticity—the ability of connections to change—is focused only on the neurons that are currently active. If a neuron fires while its input is silent, the connection between them weakens. Since the neurons for the newest task were firing during sleep, but the neurons for the older tasks were mostly quiet, the connections from the old tasks to the new ones were selectively weakened. This allowed the old tasks to regain their own dedicated pathways. The researchers found that this process did not require complex instructions or a global command center; it emerged naturally from the simple rules of how neurons interact when one group is active and another is not. Even when they simulated a brain where only excitatory connections could change, or where the rules of biological wiring were strictly enforced, the same result occurred: the new memory was separated from the old ones, and the old memories were saved.
The study also showed that this separation is not perfect for every single memory. Memories that share the most features with the newest one are the hardest to untangle, because their inputs were active enough during sleep to prevent the weakening process from fully taking hold. This mirrors real-world experiences where very similar things are harder to distinguish than very different ones. Furthermore, the researchers tested their findings across different types of digital networks and with more complex images, and the same principle held true. The mechanism appeared to be a general consequence of how local learning rules interact with biased reactivation, rather than a trick specific to one type of computer model.
These findings offer a concrete, testable explanation for a long-standing mystery in neuroscience. They suggest that REM sleep does not necessarily replay specific events to strengthen them, but rather uses the dominance of recent memories to selectively weaken the connections that cause interference. By pruning the inputs that the newest memory has captured from the past, the brain allows older memories to stand on their own again. This process ensures that as we learn new things, we do not lose the old, but rather keep them distinct and accessible. The study provides a clear blueprint for how a system can learn continuously without forgetting, offering insights that could eventually help improve artificial intelligence and deepen our understanding of how the human brain manages the constant flow of new information.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.