TL;DR
This paper proposes that "representational drift"—the constant shifting of neural activity patterns 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 a stable readout of information over time.
What Problem Does This Paper Address?
In neuroscience, a persistent mystery is why the "neural code" is so unstable. Even when an animal performs the exact same task perfectly every day, the specific neurons firing to represent that task change over time. This phenomenon is called representational drift.
While drift has been observed across various brain regions (like the hippocampus and sensory cortex), we don't fully understand the mechanism behind it or its functional purpose. Is it just metabolic noise, or does it serve a computational goal? The authors tackle this by investigating whether fluctuations in intrinsic excitability—the internal "readiness" of a neuron to fire—could be the engine driving this drift and whether that drift helps the brain organize memories chronologically.
Methods
The authors developed a rate-based recurrent neural network (RNN) consisting of 50 neurons with all-to-all plastic connections. The core of their approach involves two key components:
- Hebbian Plasticity: A standard "fire together, wire together" rule that allows the network to form and update neural ensembles (groups of neurons that represent a specific memory).
- Fluctuating Excitability: Unlike standard models where a neuron's threshold is fixed, the authors introduced a time-varying threshold ($\epsilon_i$) for each neuron. Every "day" in the simulation, a different subset of neurons was given a boost in excitability, simulating biological observations that excitability fluctuates over hours and days.
The team tested the model using a 4-day protocol: initial encoding on Day 1, followed by spontaneous or cue-induced reactivations on Days 2, 3, and 4. They also implemented two types of "decoders" (classifiers) to see if they could extract temporal information from the drifting patterns: a Day Decoder (to identify the absolute day) and an Ordinal Time Decoder (to identify the sequence of events).
Datasets
As a theoretical computational study, this paper does not use external experimental datasets. Instead, it relies on synthetic data generated through simulations. The authors simulated populations of $N=50$ neurons and validated their findings across different architectural variations, such as:
- Sparse connectivity: Testing if the results hold when neurons aren't all-to-all connected.
- Slope-based excitability: Changing the gain of the neuron's response rather than just its firing threshold.
Key Insights and Ideas
- Drift as a "Time-Stamping" Mechanism: The most significant insight is that drift isn't just random decay. Because new neurons are recruited into an ensemble based on their current excitability, the resulting pattern becomes a unique signature of when that memory was last active.
- The "Goldilocks Zone" of Excitability: For drift to be useful for decoding the order of events, the fluctuations in excitability must be in an intermediate range. If fluctuations are too low, there is no drift; if they are too high, the ensemble is completely replaced every day, losing the "thread" of the memory.
- Stability through Plasticity: Even though the ensemble members change (drift), a downstream "readout" neuron can still track the memory. By using its own plastic synapses to follow the "center of mass" of the drifting ensemble, the brain can maintain stable behavior despite a shifting underlying code.
- Independent Drifting: The model shows that multiple memories (e.g., Context A and Context B) can drift independently within the same network without interfering with each other.
Key Results
- Perfect Temporal Decoding: The "Day Decoder" was able to identify which day a neural pattern was recorded with near-perfect accuracy, proving the ensemble carries a strong temporal signal.
- Sequence Recovery: The "Ordinal Decoder" successfully recovered the correct sequence of reactivations (Day 1 $\rightarrow$ 2 $\rightarrow$ 3 $\rightarrow$ 4) significantly better than chance, provided excitability fluctuations were at an optimal amplitude ($E \approx 1.5$).
- Engram Independence: The rate of drift was found to be independent of the size of the initial neural ensemble (the "engram"), suggesting this is a robust property of the network's dynamics.
- Readout Quality: While the quality of memory readout slightly decreases as the drift rate increases ($R^2 = 0.198$), the readout neuron remained functional across all tested drift rates.
Limitations and Future Work
- Semantic Overlap: As noted by the reviewers, the model's "intrinsic excitability" term is mathematically similar to an external "input." Future work could better differentiate between truly internal fluctuations and changes in external sensory drive.
- Network Scale: The simulations used a relatively small number of neurons ($N=50$). While the authors argue the results are scale-invariant, testing this in much larger, more complex networks would be a logical next step.
- Experimental Validation: The study is purely theoretical. The authors suggest that future experiments could test this by artificially manipulating excitability (e.g., via drugs or exercise) and measuring whether the rate of representational drift changes accordingly.
Summary updated