Decomposing representational drift across wake and sleep

By tracking representational drift in the mouse olfactory cortex across wake and sleep cycles, this study reveals that sleep initiates a distinct reorganization of odor representations through decorrelation and rotation, rather than simply continuing online learning, and provides the first evidence for temporally compressed olfactory replay associated with piriform cortex sharp waves.

The Gist

Imagine your brain is like a massive, bustling library. Every time you learn something new—like the smell of a fresh coffee or a specific flower—your brain writes a "book" about it. But here's the strange thing: the librarians (neurons) who write these books don't stay the same. Over time, different librarians take over the same stories. This shifting of staff is called "representational drift."

For a long time, scientists knew this drift happened, but they didn't know why or how. Was it just the brain getting messy? Was it learning? And what role does sleep play in all of this?

This new study by Harris, Schaefer, and Kollo acts like a high-tech time-lapse camera, watching the library in real-time to see what happens when the lights go out (sleep) versus when they are on (wakefulness).

Here is the story of their discovery, broken down into simple concepts:

1. The Experiment: Smelling, Sleeping, and Smelling Again

The researchers put mice in a special chair where they could record the activity of hundreds of individual neurons in the olfactory cortex (the brain's smell center).

• Phase 1 (Wake): They let the mice smell a specific sequence of scents (like a story: "Rose, then Lemon, then Mint, then Cinnamon").
• Phase 2 (Sleep): They let the mice take a nap. No smells were presented.
• Phase 3 (Wake): They woke the mice up and let them smell the same sequence again.

They wanted to see how the brain's "code" for these smells changed between the first and second smelling session.

2. The Big Discovery: Two Different Types of Drift

The team built a clever computer program (a "decoder") to track how the brain's code changed. They found that the drift wasn't just one slow, messy slide. Instead, it was made of four distinct steps, like a dance with specific moves.

• The "Wake" Move (Adaptation): When the mice first smelled the scents, the brain's code changed slightly to get used to the smell. It was like the librarians getting comfortable with the books. The smells became a bit more similar to each other (less distinct), which is normal when you are just getting used to something.
• The "Sleep" Move (The Great Turnaround): This was the surprise. When the mice fell asleep, the brain didn't just keep drifting in the same direction. It did a 180-degree turn.
  • Rotation: The brain rearranged which neurons were telling the story. It was like swapping the entire staff of the library, but keeping the story exactly the same.
  • Decorrelation: This is the most important part. During sleep, the brain made the different smells more distinct from each other. If "Rose" and "Lemon" were starting to look a bit alike in the brain's code, sleep pushed them apart again. It's like the librarian taking two similar books off the shelf and moving them to opposite ends of the room so they never get confused.

The Metaphor: Imagine you are drawing a map of a city while walking (Wake). You might make a few small errors or shortcuts. But when you go home and sleep, your brain doesn't just keep drawing the same map. It takes a red pen, erases the confusing parts, rotates the map slightly, and redraws the streets so the "Coffee Shop" and the "Bakery" are clearly separated. Sleep isn't just resting; it's reorganizing the library for better efficiency.

3. The "Ghost" in the Machine: Olfactory Replay

The researchers also found something magical happening during sleep: Replay.

Just as we know the brain "replays" the day's events in the hippocampus (the memory center for space), they found the smell center was doing it too.

• While the mice were asleep, their neurons fired in the exact same order as the smells they had smelled earlier (Rose -> Lemon -> Mint -> Cinnamon).
• The Speed Up: This replay happened 2.5 times faster than real life. It was like watching a movie in fast-forward.
• The Trigger: These replays were triggered by tiny electrical sparks in the brain called Sharp Waves. Think of these as the "conductors" of the orchestra, telling the neurons, "Okay, let's run through the song again, but faster!"

4. Why Does This Matter?

This study changes how we think about sleep and learning.

• Sleep isn't just a pause button. It's an active construction phase. While you are awake, you are gathering data. While you sleep, your brain is actively editing, sorting, and sharpening that data so you don't get confused later.
• Drift is a feature, not a bug. The fact that the brain's code changes over time isn't a sign of it breaking down. It's a sign of it constantly optimizing itself.
• Local vs. Global: The study suggests that the smell center can do this reorganizing all by itself, without needing the whole brain to coordinate. It's like a single department in a company reorganizing its files without waiting for the CEO's permission.

The Takeaway

Your brain is a living, breathing, shifting landscape. When you are awake, you are painting the picture. When you sleep, your brain is the editor, rotating the canvas, sharpening the colors, and making sure every detail stands out clearly. This study shows us that sleep is the secret ingredient that turns a messy sketch into a masterpiece.

Technical Summary

1. Problem Statement

Neural representations of sensory information are not static; they undergo "representational drift" over time, where the specific neurons encoding a stimulus change identity. While this phenomenon is well-documented over weeks to months in the olfactory cortex, the underlying mechanisms and the relative contributions of online experience (wake) versus offline states (sleep) remain unclear.

• Key Question: Is representational drift a continuous process driven by waking experience, or does sleep initiate distinct, qualitatively different transformations?
• Gap: Previous studies have not successfully decomposed drift into components operating on different timescales (seconds to minutes) to distinguish between adaptation, consolidation, and homeostatic reorganization.

2. Methodology

Experimental Paradigm

• Subjects: Male C57BL6/J mice ($N=4$).
• Setup: Head-fixed mice equipped with chronic EEG/EMG implants for sleep staging and acute Neuropixels 2.0 probes inserted into the piriform cortex (primary olfactory cortex).
• Protocol:
  1. Wake Phase 1: Presentation of 4-element odor sequences (familiar and novel) in blocks.
  2. Sleep Phase: Natural sleep period (~60–90 mins) with no odor stimulation.
  3. Wake Phase 2: Re-presentation of the same odor sequences.
• Data Recorded: Single-unit spiking activity, local field potentials (LFP), EEG, EMG, and respiration.

Computational Approach: Low-Rank Drift Decoder

To track representational changes, the authors developed a novel time-varying linear decoder with a specific architecture:

• Model: $W(\tau) = W_0 + \sum_{i=1}^{M} \alpha_i(\tau) \cdot \Delta W_i$
  • $W_0$: Baseline weight matrix (initial state).
  • $\Delta W_i$: Low-rank perturbation matrices representing distinct drift modes.
  • $\alpha_i(\tau)$: Normalized sigmoid activation functions that learn the timing and rate of each mode's activation during the session.
• Optimization: The model was trained to decode odor identity while minimizing redundancy between modes (orthogonality constraints) and enforcing smooth temporal transitions.
• Analysis:
  • Decomposition: Identified four orthogonal drift modes (D1–D4) operating on distinct timescales.
  • Metrics: Calculated Gain (alignment with baseline), Rotation (angle of weight vector change), and Decorrelation (change in pairwise correlation between odor representations).
  • Replay Detection: Used Temporally Delayed Linear Modeling (TDLM) to detect sequential reactivation of odor patterns during sleep, testing for "sequenceness" at various temporal compression factors.
  • Sharp Wave Association: Correlated replay events with local field potential (LFP) sharp waves (SPWs) in the piriform cortex.

3. Key Contributions

1. Decomposition of Drift: First demonstration that representational drift can be mathematically decomposed into four orthogonal modes operating on different timescales within a single behavioral session (wake-sleep-wake).
2. State-Dependent Transformations: Identification that sleep and wake drive qualitatively different transformations in neural codes. Sleep does not merely continue wake drift but initiates an "about-turn" in the drift trajectory.
3. Discovery of Olfactory Replay: First evidence of olfactory replay in the piriform cortex, occurring at ~2.5× temporal compression during NREM sleep.
4. Local Mechanism: Demonstration that olfactory replay is associated with locally generated piriform sharp waves (SPWs), independent of hippocampal coordination.

4. Key Results

A. Four Orthogonal Drift Modes

The analysis revealed four distinct modes (D1–D4) with unique temporal profiles and functional effects:

• D1 (Initial Wake): Active during initial odor exposure. Characterized by gain scaling (aligned with baseline weights) and increased correlation between odor representations (adaptation/redundancy).
• D2 & D3 (Sleep Transition & Sleep): Active during the transition to sleep and NREM sleep. These modes are orthogonal to the baseline. They drive a rotation of the neural code and significant decorrelation (reducing overlap between odor representations).
• D4 (Post-Sleep Wake): Active after waking. Contributes minimally to further drift, suggesting the major reorganization occurs during sleep.

B. The "About-Turn" Trajectory

• Visualizing the drift trajectory in weight space (via SVD) showed that wake-associated drift proceeds along a dominant axis.
• Upon entering sleep, the trajectory bends orthogonally, driven by D2 and D3.
• Post-sleep, the trajectory nearly stalls, indicating that sleep-induced reorganization accounts for the majority of the session's total drift.

C. Olfactory Replay and Compression

• Replay Detection: Spontaneous activity during NREM sleep recapitulated the order of awake-presented odor sequences.
• Compression: Sequenceness was maximal at a 2.5× temporal compression (replay occurring in ~300–500 ms vs. 1 s during wake).
• Structure: Replay events included both partial (3-element) and complete (4-element) sequences.

D. Association with Piriform Sharp Waves (SPWs)

• SPW Characteristics: Large negative deflections (5–20 Hz) in the piriform LFP, occurring more frequently during NREM sleep. Unlike hippocampal SWRs, these were locally generated but showed increased high-frequency power (gamma and ripple bands) immediately following the event.
• Coupling: Replay events were significantly more likely to occur within 500 ms following a piriform SPW, suggesting SPWs act as a trigger for local consolidation processes.

5. Significance and Implications

• Mechanism of Memory Consolidation: The findings challenge the view that sleep is a passive continuation of waking learning. Instead, sleep actively reorganizes sensory representations by rotating the neural code and decorrelating inputs. This aligns with efficient coding theory, suggesting sleep optimizes the population code for discriminability and storage capacity while minimizing interference.
• Local vs. Global Consolidation: The discovery of replay coupled with local piriform sharp waves suggests that sensory consolidation can occur independently of the hippocampus, challenging the strict "hippocampal-neocortical dialogue" model for all memory types.
• Timescale Integration: The study bridges the gap between fast biological processes (synaptic plasticity, spine turnover) and slow representational drift (weeks/months), showing that the latter is the accumulation of state-dependent, short-timescale transformations.
• Future Directions: The framework provides a tool for causal intervention (e.g., closed-loop manipulation of SPWs) to test how specific drift modes contribute to long-term memory stability and behavioral performance.

In summary, this paper establishes that representational drift is a multi-component process where sleep acts as a distinct, active phase of reorganization characterized by rotation and decorrelation, driven by local replay events triggered by piriform sharp waves.