When a rat walks from one room into another, the neurons in its hippocampus that signal "you are here" (place cells) reorganize completely. A cell that fired in one corner of the first room may go silent, or fire somewhere unrelated in the second room. This is called global remapping. Slow feature analysis (SFA) is an unsupervised method that finds the functions of a sensory stream that change most slowly over time. Applied to video from a simulated walking agent, it is known to produce place cells, but only ever within a single room. This note argues that global remapping does not need a separate context signal, a clustering module, or shuffled grid inputs. It follows from the slowness objective itself whenever the agent's experience has a particular temporal structure: long stays inside each environment and rare crossings between them. The rate of those crossings, not how different the rooms look, then predicts how completely the maps decorrelate.
The argument rests on a known equivalence. On long temporal streams, SFA computes the same thing as Laplacian Eigenmaps, a spectral method that finds the smoothest functions on a graph, where here the graph links frames that follow each other in time (On the Relation of Slow Feature Analysis and Laplacian Eigenmaps). Rooms connected through a rarely used doorway form an almost disconnected graph. For such a graph, the slowest functions are either indicators of which room the agent is in, or spatial patterns confined mostly to one room. After a sparsification step, those confined patterns become place cells whose fields in one room say nothing about their fields in another. That is exactly the orthogonal, decorrelated population geometry that normative models obtain only by injecting a hand-made context variable. We propose to derive this regime, measure where it begins and ends, and test it against simpler explanations in visually rendered multi-room environments.
Core Claim
Linear SFA solves the generalized eigenproblem $A w = \lambda B w$. Here $A$ is the covariance of the input's time derivatives, $B$ is the covariance of the input itself, and the slowest features are the eigenvectors with the smallest eigenvalues $\lambda$.
Consider two rooms of size $L$, traversed at speed $v$, with a per-step probability $\varepsilon$ of crossing between them. The room-indicator function changes only at crossings, so its slowness cost scales with $\varepsilon$. The slowest within-room harmonic scales with $(v/L)^2$; this follows from the standing-wave form of SFA solutions in bounded enclosures derived in Slowness and Sparseness Lead to Place, Head-Direction, and Spatial-View Cells.
Within-room harmonics of the two rooms have nearly equal costs. Weak coupling therefore lets them mix in combinations that are aligned only near the doorway. From this we predict that the correlation between a unit's rate maps in the two rooms is controlled by the dimensionless ratio
$$ \rho = \frac{\varepsilon}{(v/L)^2}, $$
the coupling strength relative to within-room eigenvalue spacing. Small $\rho$ should give global remapping: near-zero rate-map correlation, a mix of cells silent in one room and cells with unrelated fields in both. Large $\rho$ should give a single continuous chart spanning both rooms.
The claim that is new relative to the file is threefold. First, global remapping should arise without any context input. Second, the degree of remapping should be a graded function of crossing statistics while the visual input is held fixed. Third, the failure SFA showed in the two-room benchmark (Robot Navigation using Reinforcement Learning and Slow Feature Analysis) should be reinterpreted. That study looked for one global coordinate system and treated the corridor as an obstacle. We read the corridor as the trigger that makes SFA allocate separate charts, which a downstream controller must read out per chart.
This also answers the tension the file names in Gap 4.1. Minimizing squared time derivatives does not forbid abrupt jumps. It only makes them expensive in proportion to how often they happen, so rare jumps at doorways are cheap.
Research Program
The work has three parts.
The first is analytic. We will compute the SFA/Laplacian spectrum for two and three coupled rectangular rooms under random-walk statistics. From it we will derive the predicted curve of cross-room rate-map correlation against $\rho$, and the ordering in which room-indicator features and spatial features appear among the slowest outputs.
The second is a simulation of the visual model. We use the hierarchical SFA network followed by sparse coding from Slowness and Sparseness Lead to Place, Head-Direction, and Spatial-View Cells, unchanged, trained on rendered first-person video. Two settings are used. One is a reconstruction of the Egyptian-textured two-room corridor layout. The other is a set of multi-room apartments built from ReplicaCAD layouts in Habitat 2.0 via facebookresearch/habitat-lab.
The key manipulation is decoupling crossing frequency from appearance. Each rendered room pair is kept fixed while the trajectory generator varies dwell times over more than two orders of magnitude. The visual statistics of each room are also varied, from fully distinct textures to largely shared ones.
The third part concerns local learning rules and downstream use. We replace batch SFA with the local plasticity rule of A Biologically Plausible Neural Network for Slow Feature Analysis and with Incremental Slow Feature Analysis. This tests whether remapping emerges online, and whether it emerges gradually with experience, as the clustering account of Hippocampal remapping as hidden state inference predicts behaviorally. We then reuse least-squares policy iteration on the chart-structured features, as in Böhmer's setup.
Analysis uses the remapping vocabulary of Three types of remapping with linear decoders. We ask whether SFA-produced remapping is of the encoder-change type (randomized spatial correlations) rather than the mixed-selective type (structured correlations preserved across rooms). We also test whether the room manifolds are related by approximately orthogonal transforms, the signature Learning Place Cell Representations and Context-Dependent Remapping obtains from an explicit scalar context.
Baselines
The naive reference is the trivial case: PCA on the same visual features, followed by the same sparse coding. This shows what any variance-based compression produces with no temporal objective. The file reports that variance-based methods tend to latch onto rapid pixel fluctuations, but gives no number on our setting. Beating it means near-zero cross-room rate-map correlation together with low within-room position decoding error, where PCA is expected to fail on the second.
The strongest prior SFA result is Böhmer's kernel SFA with policy iteration. It reached about 80% success on real-robot single-room test trials and needed only 2 to 3 more steps than an optimal policy in simulation. In the two-room benchmark it failed to learn a functional policy, and no number is reported. We rerun it on our two-room reconstruction with identical frames and a matched kernel and feature budget. It counts as beaten if chart-aware readout of our features yields a working cross-corridor policy and approaches the single-room success level.
The strongest remapping model in the file is the normative similarity-matching network of Learning Place Cell Representations and Context-Dependent Remapping. We give it what it requires: true coordinates and the true room label as context. It serves as an oracle for remapping geometry, not a competitor on unsupervised learning. Beating it is not the goal. Matching its decorrelation and orthogonality statistics without labels is.
The task-driven reference is the supervised LSTM navigation network of Training recurrent networks to generate hypotheses about how the brain solves hard navigation problems. That network matched a 1,000-particle filter in single environments and a 10,000-particle filter across 100 environments, but used supervised coordinates and 250,000 trials. We train it on our rendered trajectories with the same frame budget and report its remapping statistics.
The ablation that isolates the key ingredient is SFA trained on the identical set of frames with the temporal order shuffled within short windows. Shuffling destroys slowness inside each room. A second ablation, more important, keeps the frames fixed while varying only the crossing rate $\varepsilon$. If remapping tracks $\rho$ here, temporal statistics are the cause.
Why not just…?
Why not just cluster frames by appearance and run single-room SFA within each cluster? This is the real threat. When rooms look different, it will plausibly reproduce global remapping just as well. Appearance clustering does not depend on how often the agent crosses between rooms, whereas our account predicts a graded transition to a single shared chart as $\rho$ grows. The dwell-time sweep with fixed rendered rooms settles this. If remapping does not vary with crossing statistics, the slowness explanation adds nothing over appearance-based gating and we will say so.
Why not just do spectral clustering or Laplacian Eigenmaps on the temporal graph directly? It is the same mathematics in the asymptotic limit (On the Relation of Slow Feature Analysis and Laplacian Eigenmaps), so we include it as a nonparametric upper reference. The difference is practical and biological. SFA is a function of the current input, so it generalizes to new frames and runs with local plasticity, while eigenmaps must store the whole graph. The comparison measures how much the parametric version loses.
Why not just use Bayesian hidden-state clustering, or shuffle upstream grid phases? The clustering account (Hippocampal Remapping as Learned Clustering of Experience) runs on abstract simulated observation distributions. The grid-realignment model (A physiologically inspired model for global remapping in the hippocampus) assumes ready-made grid inputs. Neither says how pixels become a context. We view SFA's slow room-indicator outputs as a candidate sensory front end for the clustering model rather than a rival. A combined variant, with the clustering model run on SFA's slowest outputs, is included to test whether it adds anything beyond the slowness result.
Why not just train an end-to-end recurrent network? It produces remapping, but only with coordinate supervision and large trial counts (Training recurrent networks to generate hypotheses about how the brain solves hard navigation problems). Its remapping, unlike ours, comes with no prediction about crossing statistics. We report both under the same frame budget.
Decisive Experiments
The central test is the $\rho$ sweep in the rendered two-room and Habitat apartment settings, with appearance held fixed. We measure four things as functions of $\rho$: the median correlation of each unit's rate maps across rooms, the population-vector correlation, the fraction of units active in only one room, and how close the cross-room manifold mapping is to an orthogonal transform. The theory predicts a sigmoidal transition located near the predicted value of $\rho$ from the analytic spectrum. A flat curve would falsify the claim.
A second test uses partly shared textures. It asks whether appearance overlap shifts the transition point rather than replacing it. This would show that function-class coupling, from the shared visual features, and temporal coupling, from crossings, add together as the theory says.
A third test uses the local and incremental SFA rules on a single continuous stream. It asks whether decorrelation between rooms grows with cumulative experience, the experience-dependent remapping that the clustering account predicts.
The fourth test is applied. It checks whether chart-aware policy iteration on our features solves the two-room corridor task where Böhmer's configuration failed, compared on identical frames.
References
- On the Relation of Slow Feature Analysis and Laplacian Eigenmaps
- Slowness and Sparseness Lead to Place, Head-Direction, and Spatial-View Cells
- Robot Navigation using Reinforcement Learning and Slow Feature Analysis
- Habitat 2.0: Training Home Assistants to Rearrange their Habitat
- A Biologically Plausible Neural Network for Slow Feature Analysis
- Incremental Slow Feature Analysis
- Hippocampal remapping as hidden state inference
- Three types of remapping with linear decoders: A population-geometric perspective
- Learning Place Cell Representations and Context-Dependent Remapping
- Training recurrent networks to generate hypotheses about how the brain solves hard navigation problems
- Hippocampal Remapping as Learned Clustering of Experience
- A physiologically inspired model for global remapping in the hippocampus