Topographic Independent Component Analysis
Model Families AIMech:TopographicICA DRAFT
Aliases: Topographic ICA, TICA
A linear latent-component model whose otherwise conditionally independent source components share random scales through fixed overlapping neighborhoods, allowing nearby component energies to be dependent while supporting a topographic organization of the representation.
Identity Notes & Concept Scope
Scoped to Hyvarinen, Hoyer and Inki’s 2001 generative construction and its approximate estimation. Topography describes higher-order statistical dependence of component amplitudes, not Euclidean proximity of prototypes, causal structure or arbitrary spatial data. This differs from ordinary mutually independent ICA, Kohonen maps and the undirected Product of Student-t analysis density. Overlapping neighborhoods distinguish it from disjoint independent subspaces; the source likelihood surrogate is not silently substituted for the exact directed model. Four candidate identities were compared against all 442 pinned AIO IDs, labels and scoped synonyms and 450 preceding native IDs, names and aliases, with no normalized match. Ignored/hidden full-checkout searches returned the same 106 matched-line set in each checkout after the prior merge. All matched excerpts from 76 files were read; they are discovery context, not existing concept records. The semantic search covered all pinned labels and scoped synonyms and returned 212 rows: 205 byte-equivalent prior readings and seven additional rows, all inspected. Twenty-three annotated AIO nodes with outgoing edges comprise 21 unchanged prior readings and two new inspections. Thirty contextual YAMLs, including seven native neighbors, comprise 25 unchanged prior readings and five complete new readings. This absence assessment is bounded to the pinned snapshot and inspected maintained corpus.
Claims & Grounded Evidence
4 Claims
Mechanism
#shared-scale-model
Generate independent higher-order variables u_k and independent zero-mean unit-variance z_i, all mutually independent. For a fixed nonnegative neighborhood h, set sigma_i=phi(sum_k h(i,k)u_k), s_i=sigma_i*z_i and x=A s with square invertible mixing A, proper latent distributions and finite positive scales. Distinct s_i are uncorrelated when the needed moments exist, but their squared-value covariance equals Cov(sigma_i^2,sigma_j^2). Disjoint scale neighborhoods imply independent components under the model. Shared neighborhoods can yield energy dependence; constant scales need not. Gaussian z gives Gaussian scale mixtures, with excess fourth cumulant 3 Var(sigma_i^2) when finite. Delta neighborhoods recover independent source components.
Mechanism
#likelihood-surrogate
The exact marginal integrates the latent scales. For Gaussian z and phi(v)=v^(-1/2), the paper replaces sqrt(sum_k h(i,k)u_k) by sqrt(h(i,i)u_i) in the integrand. With nonnegative u,h and positive diagonal h this produces a pointwise lower bound, factoring as exp(sum_k G(sum_i h(i,k)s_i^2)), where G is the one-dimensional log integral in Equation 3.9. This is generally an unnormalized surrogate rather than the original marginal. Normalizing the surrogate does not preserve its pointwise lower-bound property. The later smoothed-square-root and polynomial contrast choices are not automatically exact lower bounds for a specified latent prior.
Mechanism
#approximate-fitting
After centering and whitening, optionally in a PCA-reduced subspace, constrain the square separating matrix W to be orthogonal. Write a_i=w_i^T z, local energies e_k=sum_j h(k,j)a_j^2 and r_i=sum_k h(i,k)Gprime(e_k) for symmetric h. The source contrast gradient is proportional to E[z*a_i*r_i], with a common factor two absorbed into step size. Reorthogonalize using W <- (W W^T)^(-1/2) W when full rank. On this orthogonal domain log|det W| is constant, explaining its omission. The rule differentiates the approximate contrast, not the generally intractable exact likelihood.
Limitation
#model-and-evaluation-limits
The image experiment finds energy covariance that remains above zero at long distances, indicating mismatch to the model’s distant-independence ideal. Synthetic recovery with matched topology does not prove consistency for arbitrary contrasts; the quadratic/kurtosis contrast gives unsatisfactory image results. Proper scale mixtures need not have finite variance: with delta neighborhoods and squared-standard-Gaussian u, the inverse-square-root construction gives a Cauchy component. Thus moment and whitening arguments need conditions beyond a proper probability distribution. A positive quadratic G also defines an optimization contrast, not an integrable exp(G) source density over all real coordinates.
⚠️ Curated Limitations & Negative Results
- Neighborhoods and latent hyperparameters are specified by the model; the paper does not infer an arbitrary topology. Neighborhood overlap, sufficient moments and actual data fit govern statistical interpretations.
- A proper generative marginal, its lower-bound surrogate and heuristic contrasts are different objects. For fixed complete invertible transforms a surrogate normalizer can be independent of W without making the surrogate the original likelihood.
- Whitening and moment claims require finite moments and suitable covariance rank. Orthogonalization requires a full-rank intermediate matrix. PCA reduction retains a subspace; it does not reconstruct discarded dimensions.
- Finite-step projected gradient updates do not establish monotone ascent, unique estimates or global optimization. The finite calculations check local derivatives, mixture normalization and energy covariance, not statistical consistency.
- Natural-image features and a two-minute MEG artifact illustration provide bounded applications. Complex-cell and wiring interpretations are hypotheses, not causal biological validation; adjacency is not automatically meaningful for every recovered signal.
- No author implementation, image-model training, MEG processing or benchmark was executed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | The source-defined latent-variable distribution is a specific mathematical representation within the pinned Model class. |
AIO:UnsupervisedLearning
Unsupervised Learning
|
related_match | The model is estimated from unlabeled observations to expose latent component structure. The target is a task, so this mapping does not assert model/task equivalence. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Aapo Hyvärinen, Patrik O. Hoyer, Mika Inki
|
primary paper
full_text
|
ID: doi:10.1162/089976601750264992
sha256:662379f40bc4…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T01:49:01Z
Updated: 2026-10-12T01:49:01Z
Source File:
data/curated/TopographicICA.yaml