TL;DR
This paper proposes that "representational drift"—the phenomenon where neural activity patterns change over time despite stable behavior—is driven by slow, natural fluctuations in how easily individual neurons fire (intrinsic excitability). Using a recurrent neural network model, the authors demonstrate that these fluctuations allow the brain to "time-stamp" memories while maintaining the ability to decode them accurately over several days.
What Problem Does This Paper Address?
In neuroscience, a long-standing mystery is why the brain’s internal "code" for a specific memory or task changes over time, even when the resulting behavior remains perfectly stable. This is known as representational drift. If you learn a task today, the specific ensemble of neurons firing in your hippocampus might look very different a week from now, yet you still remember how to do the task.
The authors tackle two main questions:
- What is the engine of this drift? They hypothesize that it is driven by fluctuations in "intrinsic excitability"—a neuron's internal "readiness" to fire.
- Is drift useful? Rather than being mere noise, the authors investigate whether drift provides a temporal structure that helps the brain organize memories chronologically.
Methods
The researchers developed a rate-based recurrent neural network (RNN). Unlike standard static models, this network featured:
- Hebbian Plasticity: Synapses between neurons strengthen or weaken based on co-activity ("neurons that fire together, wire together").
- Fluctuating Intrinsic Excitability: Each neuron was assigned a time-varying excitability term ($\epsilon_i$). This term changed the neuron's firing threshold or the slope of its activation function over a simulated 4-day period.
- Reactivation Protocol: The network was "stimulated" on Day 1 to encode a memory. On subsequent days, the memory was reactivated (simulating recall or sleep-based replay).
- Decoding Analysis: The authors built two types of classifiers: a Day Decoder to identify which day a pattern was recorded, and an Ordinal Time Decoder to see if the network could reconstruct the sequence of past events.
Datasets
As this is a theoretical and computational study, no external biological datasets were used. Instead, the authors generated simulated neural activity data from their RNN. They tested the model's robustness by varying parameters like network sparsity (comparing all-to-all connectivity vs. 50% sparse) and the amplitude of excitability fluctuations ($E$).
Key Insights and Ideas
- Excitability as a "Motor" for Change: When a memory is reactivated, neurons with high current excitability are "recruited" into the ensemble, while less excitable neurons drop out. Because the synapses are plastic, the network's weights remodel to favor this new, slightly shifted group.
- The "Goldilocks Zone" of Drift: The authors found that for drift to be useful, it must be gradual. If excitability fluctuations are too low ($E=0$), there is no drift and no temporal information. If they are too high ($E=3$), the memory is overwritten by noise. At intermediate levels, the drift is "just right"—preserving the memory while adding a temporal "time-stamp."
- Drift is Informative: Because the ensemble changes predictably over time, the current state of a neural representation contains a signature of its history. This allows the brain to distinguish between a memory formed yesterday and one formed three days ago.
- Self-Healing Readouts: The study shows that an "output neuron" (representing a downstream brain area) can use its own plastic synapses to "track" the drifting ensemble. As long as the change is gradual, the output neuron stays tuned to the memory, explaining why behavior remains stable despite internal flux.
Key Results
- Perfect Day Identification: The Day Decoder was able to perfectly identify the reactivation day of a neural pattern compared to shuffled data.
- Sequence Reconstruction: At intermediate excitability levels ($E=1.5$), the Ordinal Time Decoder correctly inferred the order of reactivations, proving the network encodes a chronological history.
- Independent Drift: The model successfully maintained two distinct memories simultaneously. Each drifted independently without interfering with the other, provided they were activated in different "contexts."
- Scale Invariance: Interestingly, the drift rate was found to be independent of the initial size of the neural ensemble (the "engram"), suggesting this is a fundamental property of the network's dynamics.
Limitations and Future Work
- Input vs. Excitability: As noted in the peer review section, the mathematical implementation of "input" and "excitability" in this specific model is quite similar. Future work is needed to distinguish these two biologically distinct signals more clearly.
- Simplified Stimuli: The model uses relatively simple, global inputs. Real-world sensory input is high-dimensional and continuous; testing how this model handles complex, overlapping stimuli would be a valuable next step.
- Experimental Validation: While the model aligns with observed data, the authors suggest that future experiments should attempt to directly manipulate neuronal excitability (e.g., via optogenetics) to see if it predictably speeds up or slows down the rate of representational drift in living animals.
Summary updated