---
kind: concept
slug: homology-theory
title: Homology theory in vision
summary: >-
  Homology counts the holes of each dimension in a space. One perspective proposes that visual perception splits by the parity of these counts: even dimensions as static scaffolds that bind features, odd dimensions as flows that carry change. Covered: one paper, a hypothesis with no experiments and no algorithm. The trade-off: it offers vocabulary, not a method to implement.
tags: [topology, visual-perception, binding, invariance, neuroscience]
papers: ["1802.04723"]
related: [persistent-homology, invariant-learning, manifold-alignment]
status: draft
updated_at: 2026-11-22
---

## Overview

Homology groups H_k = Z_k / B_k record the k-dimensional holes of a space. Z_k
are the cycles, B_k the boundaries, and the Betti numbers b_k = dim H_k count
the independent holes [arxiv:1802.04723 § Hypothesis] (From Consistency to
Conservation). The boundary operator satisfies ∂² = 0, and the Euler-Poincaré
identity links cell counts to Betti numbers [arxiv:1802.04723 § Hypothesis]
(From Consistency to Conservation).

The paper's use of this theory is a claim about perception. Even-dimensional
classes (b_0, b_2, ...) are static scaffolds that hold structure or context.
Odd-dimensional classes (b_1, b_3, ...) are dynamic flows that carry content or
process [arxiv:1802.04723 § Hypothesis] (The Parity Hypothesis). The author
argues that one such split covers three problems that are usually studied
apart: binding, object invariance and the inverse problem [arxiv:1802.04723
§ Introduction] [arxiv:1802.04723 § Applying].

Only one card on wikillm lists this concept, so this page describes one
perspective. The paper has no experiments and no algorithm with defined
operators, so every claim below is a hypothesis.

## Variants

### Parity split of scaffolds and flows

The single framework the paper offers splits a latent representation by the
parity of its homology. Its parts are a topological premise, a parity
hypothesis, a neural mapping, an objective and a set of operators
[arxiv:1802.04723 § Hypothesis] [arxiv:1802.04723 § Objective]. Card:
[The Shape of Sight](/papers/1802.04723).

The objective is stated as minimising content-context uncertainty. It is said
to be minimal exactly when the model is topologically self-consistent, with
χ ≈ 0 [arxiv:1802.04723 § Objective] (The Duality). The paper does not derive
this equivalence, and it does not define the learning rule that reaches the
fixed point [arxiv:1802.04723 § Objective] (The Duality).

The applications, each stated as a hypothesis, are:

- Binding: features count as bound once they jointly drive one two-dimensional
  scaffold [arxiv:1802.04723 § Applying] (Binding is Even-Dimensional
  Scaffolding).
- Invariance: separating an object's identity from its transformations is read
  as a parity separation of scaffolds from flows [arxiv:1802.04723 § Applying]
  (Manifold Unfolding is Parity Sorting).
- Multistable perception: switching between two stable percepts is a traversal
  along a one-dimensional loop that joins two scaffolds [arxiv:1802.04723
  § Applying] (Resolving Ambiguity).
- Inverse problem: the 2D input selects the best-fitting 3D template from a
  pre-built library, which the paper calls verification rather than search
  [arxiv:1802.04723 § Applying] (The Inverse Problem as Template Selection).

## When to use which

No variant can be implemented from the paper alone, so the table compares ways
of using the idea, not competing methods.

| Situation | Use | Why |
|---|---|---|
| You want a vocabulary for why binding and invariance are hard | The parity split as a conceptual lens | The paper names both problems and proposes one split for them [arxiv:1802.04723 § Introduction] |
| You need a training or inference algorithm | Not this paper | The operators F and R are named but not defined, and no stopping rule is given [arxiv:1802.04723 § Objective] (The Functional Split) |
| You need to compute holes from data | See the [persistent homology](/concepts/persistent-homology) page | That page covers a method with an implementation, from a different paper; the two do not share a result here |
| You want to test the claim empirically | Start from the paper's stated predictions, but first check they are in the published text | The paper runs no experiments, and its testable predictions sit in a block the source comments out (see Pitfalls) |

## Implementation sketch

The paper gives no algorithm. The one computable piece is the Euler-Poincaré
identity, which relates cell counts c_k to Betti numbers b_k. The source does
not say which complex a learned representation would use, or how the Betti
numbers would be obtained, so the sketch below takes the counts as given.

```pseudocode
# Check the Euler-Poincaré identity for a finite cell complex
# [§ Hypothesis] (From Consistency to Conservation)
# Input:  cell_counts[k] = c_k, number of k-cells
#         betti[k] = b_k, Betti numbers (source: not stated)
# Output: chi, and whether the two sides agree

def euler_check(cell_counts, betti):
    # Alternating sum over cells: chi = sum_k (-1)^k c_k.
    chi_cells = sum((-1) ** k * c for k, c in enumerate(cell_counts))
    # Alternating sum over Betti numbers: sum_k (-1)^k b_k.
    chi_betti = sum((-1) ** k * b for k, b in enumerate(betti))
    # The paper's objective needs chi near 0 [§ Objective] (The Duality);
    # the tolerance is not stated.
    return chi_cells, chi_betti, chi_cells == chi_betti
```

Where the variant differs: the paper's parity split would assign the even
entries of `betti` to scaffolds and the odd entries to flows [§ Hypothesis]
(The Parity Hypothesis). It gives no way to compute those entries from a
network or from neural data.

## Contradictions

- Content and context are swapped in one place. The text defines Φ as content
  and Ψ as context [arxiv:1802.04723 § Objective] (The Duality). Figure 3
  labels Φ_t as context, on H_odd, and Ψ_t as content, on H_even [arxiv:1802.04723
  fig. 3]. The rate-code section assigns content (Φ) to H_even [arxiv:1802.04723
  § Applying] (The Parity of Neural Codes), while the functional split assigns Φ
  to the H_odd flow [arxiv:1802.04723 § Objective] (The Functional Split). The
  paper does not resolve this, so pick one convention before using these labels.

## Papers

- [The Shape of Sight](/papers/1802.04723):
  a perspective that splits visual perception by the parity of homology. It
  proposes the scaffold-and-flow split for binding, invariance and the inverse
  problem, with no experiments and no algorithm.
