Lateral-Interaction Smoothing¶

Local smoothing method — instantiates Neighborhood-Preserving Substrate Mapping

Couples each substrate unit to its immediate neighbors — near-excite, surround-suppress — so adjacent units come to represent similar inputs, giving the map graded, overlapping receptive fields and local continuity.

Lateral-Interaction Smoothing wires each substrate unit to its spatial neighbors with a short-range excite / longer-range inhibit profile — a "Mexican-hat" kernel — so that when one unit responds, its near neighbors are nudged to respond similarly and its distant ones are suppressed. Its defining move is enforcing local continuity from the bottom up: it does not lay out the whole map (a construction method does that) — it shapes the fine texture, making neighboring units' receptive fields overlap and vary smoothly rather than jump. That is what turns a set of independent detectors into a continuous map, and the overlap it creates is deliberate redundancy: adjacent units share response, so no single unit is a point of failure.

Example¶

In primary visual cortex, neurons tuned to edge orientation are arranged so that preferred orientation rotates smoothly across the cortical surface rather than jumping randomly cell to cell. Lateral connections are a large part of why: short-range excitatory and broader inhibitory coupling means a neuron's activity reinforces its near neighbors and suppresses distant ones, pulling nearby cells toward similar tuning. The result is the smooth "pinwheels" of orientation, with receptive fields that overlap, so a small patch of cortex covers a continuous range of orientations and the loss of one cell barely dents coverage. A model builder adds the same lateral-interaction smoothing to an artificial feature map to get that graded continuity, instead of a salt-and-pepper jumble where adjacent units answer to unrelated inputs.

How it works¶

Define a neighborhood kernel over the substrate — excitatory at the center, inhibitory in the surround. During learning or settling, let each unit's response propagate through the kernel so neighbors co-adapt toward similar selectivity. The fixed point is a map whose receptive fields vary smoothly and overlap. What distinguishes it is that it operates on local unit-to-unit coupling to shape receptive-field structure and overlap — it is neither the global layout method nor a measurement of the result.

Tuning parameters¶

Excitatory radius (kernel width) — how many neighbors co-activate; wider makes a smoother, coarser map with larger receptive fields.
Surround-inhibition strength — how sharply distant units are suppressed; sets how crisp the map's feature boundaries come out.
Excite/inhibit balance — the Mexican-hat shape, tilting the operation toward blurring or toward sharpening.
Overlap fraction — how much adjacent receptive fields share; more overlap buys redundancy at the cost of selectivity.
Settling iterations — how long the lateral dynamics run before the map is read out.

When it helps, and when it misleads¶

Its strength is smooth, continuous maps with graceful degradation — because adjacent fields overlap, a single unit's loss leaves a small, filled-in dent rather than a hole. Its failure modes are the excesses of smoothing. Too wide a kernel over-smooths, destroying real discontinuities the source genuinely has by ramping a true boundary into a gradient; too much overlap wastes substrate on redundancy and dulls selectivity so nearby distinct inputs become indistinguishable. The classic misuse is smoothing a map that should carry sharp seams because a smooth picture simply looks nicer. The discipline is to match kernel width to the source's actual smoothness and to preserve genuine discontinuities rather than blur them away.^[1]

How it implements the components¶

It shapes the local receptive-field texture and the overlap band; it neither builds the global map nor scores it:

local_receptive_field_or_jurisdiction — the neighborhood kernel shapes each unit's receptive field, its size and smoothness.
overlap_and_redundancy_band — it deliberately makes adjacent fields overlap, creating the redundancy band that gives graceful degradation.

It does not lay out the global correspondence — that is Elastic-Net Embedding or Self-Organizing Map Training — and it does not define the tiling grid the fields sit on, which is the Receptive-Field Tiling Grid.

Instantiates: Neighborhood-Preserving Substrate Mapping — it supplies the local continuity and overlap that make the map genuinely topographic.
Sibling mechanisms: Receptive-Field Tiling Grid · Self-Organizing Map Training · Adaptive Remeshing and Reallocation · Adjacency-Matrix Preservation Test · Boundary and Seam Regression Test · Calibration Anchor Stimuli · Canary Region Probe · Coverage-Hole Heatmap · Elastic-Net Embedding · Local Ablation or Lesion Probe · Magnification Function · Map Fold and Collision Scan · Map Registration and Alignment · Multidimensional Scaling Layout · Neighborhood Trustworthiness and Continuity Metric · Perturbation-Response Map · Topographic Error Measure · Versioned Coordinate Atlas

References¶

[1] Lateral inhibition with a center-surround ("Mexican-hat" / Difference-of-Gaussians) profile is a well-documented motif in sensory neural maps. Cited as the real, correctly-used anchor for neighbor coupling. ↩