Homeomorphism

As explained in Math and Architectures of Deep Learning

A topology-preserving mapping between manifolds that allows stretching and squishing but not tearing, breaking, or folding.

A homeomorphism transforms one manifold into another by stretching or squishing without cutting, breaking, or folding. It preserves path-connectedness and can make geometric structures easier to analyze or classify. section 12.2 section 12.4

Formally, functions f and f⁻¹ connect sets X and Y: both are one-to-one and continuous, so nearby points remain nearby in both directions. A circle arc can therefore become a line segment, and a spherical patch can become a plane patch. section 12.2

Making classes easier to separate

Neural-network layers can successively transform an input manifold until classes that originally surround one another become separable by a line or plane. A linear transformation, translation, and sigmoid are continuous; when the weight matrix is square and has nonzero determinant, the transformation is invertible and therefore homeomorphic. A zero determinant instead causes dimensionality reduction. section 12.3

Use the geometric figure to compare stretching a circle arc into a line and mapping a spherical patch onto a plane. section 12.2

How Math and Architectures of Deep Learning covers it

How deep it goes

Defines bijectivity, continuity, inverse continuity, and path-connectedness, while remaining an intentionally non-rigorous geometric introduction.

Examples

Uses circle arcs, spherical patches, surrounded classes, curved lines, and bell-like manifolds as geometric examples.

What you can do after

Understand how continuous invertible transformations can reshape manifolds and support easier separation by a final linear classifier.

How it connects

Connects manifolds, path-connectedness, neural networks, linear layers, sigmoid nonlinearities, and linear classification.

liveBookmanning.com

Read Math and Architectures of Deep Learning on liveBook Get the book at manning.com

Related concepts

• continuous functionimplements · A function whose output changes arbitrarily little when its input changes arbitrarily little.
• manifolduses · A manifold is a mathematical space that locally resembles Euclidean space, although its global geometry or topology may be more complex.
• linear layerproduced by · A linear layer is a neural-network layer that maps input features to outputs by a learned affine transformation, typically connecting each output to every input feature.
• data transformationkind of · A data transformation is an operation or function that maps input data or its representation to modified data or a different representation.

This explainer was generated by Manning's knowledge graph from the text of Math and Architectures of Deep Learning. The section links point to the parts of the book it draws on.