When and Where: A Model Hippocampal Network Unifies Formation of Time Cells and Place Cells

This paper proposes a single recurrent neural network model of the hippocampal CA3 region that unifies the formation of place and time cells by demonstrating how the network's dynamical regime shifts between stable attractors and sequentially broadened fields depending on whether it is trained on spatial navigation or temporal sequences.

The Gist

Imagine your brain's hippocampus as a super-smart, predictive movie editor. Its job isn't just to record what's happening right now; it's to fill in the blanks when the camera cuts out or when the script gets fuzzy.

This paper argues that the brain doesn't need two different "editors" to handle where you are (Place Cells) and when things happen (Time Cells). Instead, it uses the same editor (a single network of neurons) to do both jobs, depending entirely on the type of "movie" it's trying to edit.

Here is the breakdown of their discovery using simple analogies:

1. The Two Jobs: "The Map" vs. "The Clock"

• Place Cells (The Map): These are neurons that fire when you are in a specific spot, like a street corner in New York. They tell you where you are.
• Time Cells (The Clock): These are neurons that fire in a specific sequence while you are waiting for something, like a countdown. They tell you how much time has passed.

For a long time, scientists thought these were two completely different machines in the brain. One was a "continuous map" (Place Cells), and the other was a "leaky bucket" that slowly fills up with time (Time Cells).

2. The Big Idea: One Editor, Two Scripts

The authors built a computer model of the hippocampus (specifically the CA3 area) and treated it like a predictive autoencoder.

• The Analogy: Imagine you are watching a movie, but every few seconds, the screen goes black (missing input). Your brain has to guess what happens during the black screen to keep the story going.
• The Experiment: They trained this computer brain to "fill in the blanks" in two different scenarios:
  1. The Navigation Script (Place Cells): The animal is walking around a room, but the camera is glitchy. The brain has to guess the animal's location based on partial clues.
     • Result: The brain creates a stable "map." Neurons light up like streetlights on a map, showing exactly where the animal is.
  2. The Waiting Script (Time Cells): The animal is sitting still, waiting for a bell to ring. There are no visual clues, just the passage of time.
     • Result: The brain creates a "countdown." Neurons fire one after another in a sequence, getting broader and broader as time goes on, effectively marking the seconds passing.

3. The "Sliding Scale" Discovery

The most exciting part of the paper is that these aren't two separate switches. They are on a sliding scale.

• The Metaphor: Think of the brain's activity as a dimmer switch.
  • If you give the brain mostly visual clues (walking around), the "Place Cell" light is bright, and the "Time Cell" light is dim.
  • If you take away the visual clues and make the animal wait in the dark, the "Time Cell" light gets brighter, and the "Place Cell" light fades.
  • The Magic: You can slowly turn the dial from "Walking" to "Waiting," and the neurons smoothly transform from acting like a map to acting like a clock. They don't switch gears; they just shift their focus.

4. Why Does This Happen? (The Mechanism)

The paper explains how the brain does this. It's all about filling in the missing story.

• When you are walking: The "missing story" is a gap in your location. The brain fills it by creating a stable spot (a Place Cell) to anchor the memory.
• When you are waiting: The "missing story" is a gap in time. The brain fills it by creating a sequence of events (Time Cells) to mark the duration.

The authors found that the brain changes its internal wiring (the connections between neurons) slightly to suit the task.

• For Time, the neurons connect in a "forward chain" (like a line of dominoes falling), pushing the signal forward through time.
• For Space, the neurons connect to create a "stable bump" (like a ball sitting in a valley), holding the position steady.

5. The Takeaway

The brain is incredibly efficient. It doesn't build a separate factory for "Time" and a separate factory for "Space." Instead, it has one universal prediction engine.

• If the world is full of spatial clues, the engine becomes a GPS.
• If the world is full of time gaps, the engine becomes a Stopwatch.

In short: Place cells and Time cells are just the same neurons wearing different hats, depending on whether the story they are trying to tell needs a map or a clock. This suggests that our sense of "where" and "when" are deeply intertwined, built on the same foundation of remembering and predicting our experiences.

Technical Summary

1. Problem Statement

The hippocampus contains two distinct types of neurons critical for episodic memory: place cells (which fire selectively at specific spatial locations) and time cells (which fire sequentially during time delays). While both cell types reside in the same subfield (CA3) and share a dense recurrent neural substrate, existing computational models treat them as fundamentally different mechanisms:

• Place cells are typically modeled as Continuous Attractor Networks (CANs), relying on path integration and stable activity bumps.
• Time cells are often modeled as leaky integrators or single-neuron mechanisms performing inverse Laplace transforms.

The Gap: These approaches fail to explain how the same recurrent network can generate both cell types, how they co-occur, or how they might transition into one another. The authors propose that these distinct representations are not separate constructs but rather different dynamical regimes of a single network driven by task statistics and input structure.

2. Methodology

A. The Core Model: Predictive Autoencoder

The authors model the hippocampal CA3 region as a Continuous Time Recurrent Neural Network (CTRNN) functioning as a predictive autoencoder.

• Architecture: A three-layer network (Input $\to$ Recurrent Hidden Layer $\to$ Output).
• Dynamics: The hidden state $v_t$ evolves according to membrane time constants, recurrent weights ($W_{rc}$), input weights ($W_{in}$), and noise. Firing rates are generated via a ReLU activation function.
• Objective: The network is trained to reconstruct "experience vectors" (EVs) from partially occluded (masked) and noisy inputs. This forces the network to learn internal dynamics to "fill in" missing sensory information (pattern completion).

B. Input Generation (Experience Vectors)

The model unifies spatial and temporal inputs under a single framework of high-dimensional vectors ($D=100$ channels):

1. Spatial Inputs: Generated by an agent traversing an environment (square room or circular track). Sensory channels are defined by Gaussian random fields modulated by location.
2. Temporal Inputs: Generated by discrete events (e.g., bell rings) separated by fixed intervals. Channels show activity bumps at specific times.
3. Mixed/Conjunctive Inputs: The authors create a continuum by varying the ratio of spatially modulated channels to temporally modulated channels, or by masking spatial inputs for specific durations to simulate time delays.

C. Training and Analysis

• Loss Function: The network minimizes Mean Squared Error (MSE) between reconstructed and ground-truth EVs, plus a regularization term to keep firing rates biologically plausible.
• Cell Identification:
  • Place Cells: Identified via Spatial Information Content (SIC). High SIC indicates strong spatial selectivity.
  • Time Cells: Identified by sorting neurons by peak firing time and checking for temporal broadening (a positive correlation between peak firing time and the width of the firing field).

3. Key Results

A. Emergence of Time Cells from Temporal Inputs

When trained on purely temporal sequences (two events separated by a delay where input is masked), the network spontaneously develops time cells.

• Hidden units fire sequentially between the two events.
• Neurons firing later in the interval exhibit progressively broader temporal tuning fields, recapitulating the hallmark phenomenology of biological time cells.
• This occurs without explicit time-encoding mechanisms; it emerges from the need to predict future inputs.

B. Emergence of Place Cells from Spatial Inputs

When trained on spatial traversal data (with partial masking), the network develops place cells.

• Hidden units exhibit stable, location-specific firing fields (high SIC).
• This reproduces previous findings that CA3 acts as a spatial attractor.

C. The Spacetime Continuum and Task-Driven Transitions

The most significant finding is the smooth transition between regimes based on input structure:

• Spacetime Task: When an agent runs on a circular track, the first lap provides spatial input, and the second lap is masked (requiring prediction). Neurons active in the second lap behave like predictive place cells: they retain spatial selectivity but also encode elapsed time, showing intermediate temporal broadening.
• Deformation Experiments:
  • Spatial $\to$ Temporal: Gradually reducing the duration of spatial input windows in a circular track task causes place cells to vanish and time cells to emerge.
  • Temporal $\to$ Spatial: Gradually widening temporal event windows transforms time cells into place-like representations.
• Asymmetry: The transition is asymmetric. Spatial inputs (continuous and dense) are robust; even with few spatial channels, place-like activity persists. Conversely, time-cell-like broadening only emerges strongly when spatial input is largely absent or when temporal structure dominates.

D. Mixed Input Competition

In tasks with mixed spatial and temporal channels:

• Spatial information dominates the representational space. A small fraction of spatial channels is sufficient to maintain a large population of place cells.
• Time-cell-like activity only becomes prominent when temporal channels vastly outnumber spatial ones, suggesting a competition for shared representational resources.

4. Mechanistic Interpretation

The authors analyze the recurrent connectivity motifs ($W_{rc}$) to explain how these dynamics arise:

• Idealized Time Cell Motif: They constructed a theoretical matrix with forward excitation (neurons excite later-firing neighbors) and backward inhibition (neurons inhibit earlier-firing neighbors).
• Validation:
  • Replacing the trained $W_{rc}$ with this idealized matrix (while keeping input/output weights) preserved sequential firing and temporal broadening.
  • The trained $W_{rc}$ from the time task closely resembles this idealized motif in terms of weight distribution, eigenvalue spectra (circular law with a semi-circle bias), and singular value spectra.
  • In contrast, the space-task $W_{rc}$ exhibits a square-like eigenvalue distribution, characteristic of stable attractor dynamics (CANs).
• Unified View: Both cell types arise from the network's attempt to reconstruct missing segments of a trajectory in a high-dimensional representational space.
  • If the missing segment is spatial, the network forms a stable attractor (Place Cell).
  • If the missing segment is temporal (a gap in time), the network propagates activity sequentially to bridge the gap (Time Cell).

5. Significance and Contributions

1. Unification of Mechanisms: The paper provides a unified computational framework showing that place and time cells are not distinct neural constructs but dynamical regimes of a single recurrent autoencoder, determined by the statistics of the input and the nature of the missing information.
2. Predictive Coding: It reinforces the hypothesis that the hippocampus functions as a predictive autoencoder, where memory formation is the reconstruction of incomplete sensory trajectories.
3. Task-Dependent Plasticity: It demonstrates that the functional role of CA3 neurons (spatial vs. temporal) is flexible and driven by task demands rather than fixed anatomical wiring.
4. Testable Predictions:
   • CA3 representations should transition smoothly between place and time coding as the ratio of spatial to temporal cues changes.
   • The transition should be asymmetric (spatial cues are more robust).
   • Distinct functional connectivity motifs (forward excitation vs. attractor stability) should be observable in CA3 depending on the task, despite fixed anatomy.

Conclusion: The study suggests that the brain does not need separate mechanisms for "where" and "when." Instead, a single recurrent network, tasked with predicting future experience from partial inputs, naturally generates both spatial maps and temporal sequences as solutions to the problem of reconstructing missing sensory trajectories.