Transforming autoencoder
Architectures AIMech:TransformingAutoencoder DRAFT
A deterministic neural architecture that learns first-level capsules from images, transformed target images and supplied transformations. Each capsule encodes feature presence and pose, applies the known transformation to its pose representation, and contributes a presence-gated reconstruction to the transformed output image.
Identity Notes & Concept Scope
At the actual merged cohort338 base 09cd8b8d5cc8b6d4098c149cbdea12e797468811, the 2026-10-15 current collection covers all 751 native and 442 AIO record identity fields and byte hashes, plus all 909 main and 393 nested frontier objects mechanically. Four ignored-inclusive, hidden-inclusive, case-insensitive whole-word searches retain 13654 occurrences in 1679 exact content groups. Root read all 29 non-domain candidate-bearing contents and all 19 selected current frontier objects; 1145 NLM-domain-only groups were mechanically separated, not manually read as complete bodies. The 22 NLM catalog/namespace groups refer to the National Library of Medicine, while seven candidate-bearing groups preserve existing source leads or research notes. Same-team context audit reads 35 maintained-record matching strings, 13 current definitions and matching windows in 55 long groups among the 505 context-only groups; its scope is attributed, not exhaustive manual reading. No candidate normalized identity collision was found. This is bounded current identity clearance, not ontology-wide semantic exhaustion. The original known-transform, capsule-based reconstruction architecture is distinct from the maintained learned-warp SpatialTransformer, source/geometry-preserving CompositionalSpatialTransformer, tied InverseCompositionalSpatialTransformerNetwork and supplied-camera PerspectiveTransformerNets. Resolve existing owner237 using Hinton et al. 2011; the owner's 2015 spatial-transformer URL remains discovery provenance. Retain the related AutoEncoderNetwork mapping without asserting equivalence or conflating later capsule routing.
Claims & Grounded Evidence
6 Claims
Mechanism
#known-transformation-architecture
The network takes an image and a desired transformation as inputs and predicts the corresponding transformed image. Training uses transformed input–target pairs and direct access to their transformation, without requiring advance labels for each capsule’s visual entity or the origin of its coordinate frame. The demonstrated construction learns first-level capsules; agreement among higher-level part–whole pose predictions is motivating context.
Mechanism
#translation-capsules-and-gating
In the translation construction, each capsule has its own recognition units that compute x, y and p, where the paper interprets p as feature presence. Its generation units receive x+Delta_x and y+Delta_y. The capsule’s contribution to the output image is multiplied by p, and separate capsules cooperate at the final layer. This is a deterministic feedforward computation; the presence interpretation does not itself establish probability calibration.
Mechanism
#observation-domain-training
The recognition and generation weights are learned by backpropagating the discrepancy between the actual and target output images. Transforming the encoded pose before generation trains the representation through error measured in the observation domain. The source does not require direct target coordinates for each learned part and does not specify one universal loss formula or optimization guarantee for every configuration.
Mechanism
#matrix-transform-configuration
For the two-dimensional affine configuration, each capsule emits nine real values arranged as a 3x3 matrix A. The supplied transformation matrix T acts on these outputs as T A, and the resulting entries feed the generation units. The source reports translation, rotation, scaling and shear transformations with this construction; it does not state a hard affine or rotation-manifold constraint on the nine emitted values.
Scope
#stereo-configuration
The preliminary stereo-car configuration uses local capsules whose recognition units read corresponding patches from both images of a stereo pair. Each capsule outputs a matrix representation of feature orientation and a presence value; a supplied source-to-target transformation is applied before generation. Presence-gated generation contributes to larger output patches in both stereo images. The described experiment uses two recognition layers and no weight sharing, so shared-weight convolution is not required by this architecture.
Scope
#limited-competence-and-single-instance
A capsule represents one instance of its visual entity with one pose vector over a limited domain. The source illustrates inaccurate coordinate responses when shifts exceed a capsule’s competence and qualifies their harmlessness by suppression of its presence output. Multiple nearby instances of the same entity can confound one capsule even when its domain is small. These are limits of the representation, not guarantees that training learns correct gates or handles arbitrary transformations.
⚠️ Curated Limitations & Negative Results
- The method uses known transformations and transformed target images during training. Object-class labels and per-part pose labels are not required in the described setup, but that does not remove transformation supervision. Applicability depends on useful transformation information, observable image changes, capacity and optimization; the source’s rhetoric about forcing a manipulable property into the representation is not a general identifiability theorem.
- Reconstruction alone does not prove unique semantic parts, a unique pose convention, calibrated feature-presence probabilities or exact global equivariance. At the functional level, for fixed invertible S and sufficiently flexible generation, A_prime=A S and G_prime(B)=G(B S^-1) preserve G_prime(T A_prime)=G(T A). This is a coordinate-nonuniqueness illustration, not a claim that the reported trained models fail or that no equivariant representation can be learned.
- The printed affine operation is T A. Nine unconstrained real outputs arranged as a 3x3 matrix do not by themselves enforce an affine homogeneous last row, orthogonality, determinant +1 or invertibility. Calling a learned matrix an orientation representation does not certify membership in SO(3). The matrix action and learned generator are not a complete physical renderer or an implementation specification for all image transformations.
- One capsule supplies one pose vector and can be ambiguous for simultaneous instances of its entity within its domain. Restricting domains or adding capsules can alleviate that problem but does not prove arbitrary-instance recovery. The finite-domain example depends on suppressing inappropriate capsule contributions; calibration and reliable suppression outside the trained range are not established.
- The evidence concerns trained first-level capsules and specific digit and preliminary stereo-car demonstrations. It does not establish a complete trained part–whole hierarchy, later dynamic routing, universal recognition or reconstruction accuracy, or a general runtime advantage. Broad criticism of convolutional networks and comparisons with Kalman filtering in the discussion are not adopted as universal equivalence or superiority claims.
- Evidence uses the exact eight-page Hinton author-hosted PDF and its own page locators. The separately acquired Wang author copy differs in abstract, section and figure numbering, discussion placement and references; publisher metadata does not establish publisher-PDF byte identity. Publication precision is the year 2011, distinct from PDF build metadata and actual source access. Historical inspection did not include author code, model training or benchmark reproduction. Exact optimizer settings, numerical loss conventions and implementation behavior are not certified.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:AutoEncoderNetwork
Auto Encoder Network
|
related_match | The encoder–generator learns through image reconstruction, supporting an architectural association with the pinned Auto Encoder Network class. That target additionally specifies unsupervised pretraining and regeneration of the input, whereas this method uses externally supplied transformations and changed target images. The relation preserves the reconstruction connection without asserting exact equivalence or unqualified subclass membership. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
hinton2011
|
Geoffrey E. Hinton, Alex Krizhevsky, Sida D. Wang
|
primary paper
full_text
|
ID: doi:10.1007/978-3-642-21735-7_6
sha256:e338b67f8626…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-15T11:08:12Z
Updated: 2026-10-15T11:08:12Z
Source File:
data/curated/TransformingAutoencoder.yaml