NFMap: Topographic 2D Manifold Learning with Invertible Normalizing Flows

summary

The gist

The gist NFMAP introduces an invertible, parametric dimensionality reduction method that combines Generative Topographic Mapping's fixed latent grid with the exact likelihood of a normalizing flow.

In short

NFMAP is an invertible method that combines Generative Topographic Mapping's fixed grid prior with a normalizing flow to map data into a 2D space. It creates a navigable map and an off-manifold residual, ensuring full invertibility. This allows new points to be projected onto the map while providing an exact likelihood signal for data quality.

Key concepts

Invertible Normalizing Flow: A type of mathematical transformation that is perfectly reversible. Unlike standard methods, this flow allows you to take a point in the low-dimensional space and accurately calculate its original position in the high-dimensional data space. This property is crucial for mapping new points back to real data.
Generative Topographic Mapping (GTM): A prior design where the latent map centers are fixed onto a regular 2D grid. This provides a structured, parametric embedding for the data, meaning any location on this grid can be used as a reference point in the learned low-dimensional space.
Residual Scale ($\sigma_{\perp}$): A control parameter that determines how much of the data's structure is forced into the visible 2D map versus being kept in the off-manifold residual. Decreasing this scale prioritizes global distance fidelity, pushing more geometric information into the main map coordinates.
Student-t Mixture Model (SMM): A prior component used to handle 'map crowding' in large datasets. It uses heavy-tailed distributions instead of simple Gaussian ones for the 2D map components, allowing the model to better accommodate many data points without losing structure.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 20 episodes
Retrieval-Augmented Generation (RAG) — explained in 18 episodes
Federated Learning — explained in 17 episodes
Quantum Fisher Information (QFI) — explained in 16 episodes
Flow Matching — explained in 15 episodes

This episode discusses

The paper

NFMap: Topographic 2D Manifold Learning with Invertible Normalizing Flows · Read on arXiv

Hélena Alexandra Gaspar

We present NFMap, a dimensionality reduction method for 2D visualisation that extends Generative Topographic Mapping (GTM) with a bijective normalizing flow. Like GTM, NFMap places a fixed mixture grid prior over a 2D latent map, giving a parametric embedding, an explicit likelihood, and the ability to project new points onto an existing map; the likelihood is an effective out-of-distribution signal on chemical datasets. Unlike GTM, whose latent-to-data mapping is an RBF network of bounded flexibility, NFMap's mapping is a normalizing flow, and is therefore fully invertible: any location on the map can be carried back to data space. To alleviate map crowding on large datasets, NFMap additionally supports heavy-tailed Student-t mixture components. NFMap maps data to a latent space of the same dimension, split into a 2-dimensional navigable map and an off-manifold residual. Because the flow is bijective and its objective penalises volume contraction, variance squeezed out of the residual must reappear in the navigable map. The residual scale is the main control between local neighbourhood fidelity and global distance fidelity; its effect depends on the dimensionality of the data and is coupled to the extent of the latent grid.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "NFMap: Topographic 2D Manifold Learning with Invertible Normalizing Flows".

Jane: The gist NFMAP introduces an invertible, parametric dimensionality reduction method that combines Generative Topographic Mapping's fixed latent grid with the exact likelihood of a normalizing flow.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper called NFMap: Topographic 2D Manifold Learning with Invertible Normalizing Flows <ref:2610.16750#pg1,NFMap: Topographic 2D Manifold Learning with Invertible Normalizing Flows>. What does that actually mean for us?

Jane: It’s about taking something called Generative Topographic Mapping, or GTM, and adding a normalizing flow to it. Essentially, GTM gives you a fixed grid prior over the map structure, but NFMap makes the mapping fully invertible with this flow structure.

Lu: The authors are tackling the problem of having an explicit likelihood and an invertible mapping simultaneously in a visualization setting where you want a navigable map. It’s about making sure you can project new points onto that map and then get back to the original data space exactly.

Meng: So, what's the core idea here? Is it just another way to draw a pretty picture of high-dimensional data? Because we need something that actually works when we have real chemical datasets.

Lalam: It means they are combining a fixed grid prior with the exact likelihood from a normalizing flow, which gives us an effective out-of-distribution signal on those chemical datasets, and it’s fully invertible so you can track any point back to its source.

The paper's summary: Tom: Okay, so the main point of NFMap is this split latent space. They map your data into a latent space that has the same dimension as the original data, but they split it into a two-dimensional navigable map and an off-manifold residual.

Jane: That residual is key because it lets them control how much information goes into the visible coordinates versus what stays in that off-manifold part. They use this to balance local neighborhood fidelity with global distance fidelity.

Lu: The mapping itself is built using a bijective normalizing flow, which they construct from B affine coupling blocks. This means the mapping is invertible in closed form, and the latent vector z is partitioned into z parallel and z perpendicular to that map, where z parallel is just two dimensions.

Meng: So, if I'm trying to deploy this on a production system, what does that residual part actually do? Does it complicate things when we are just trying to visualize the main structure?

Lalam: It means the entire latent vector f theta maps the original data Xtilde exactly, and you only plot z parallel while retaining z perpendicular so no information is discarded, which means f minus one theta reconstructs Xtilde exactly >

The paper's improvements: Tom: Their main contribution seems to be how they control the trade-off between local and global structure. They introduce a residual scale, sigma sub perp.

Jane: That scale is a major control mechanism. It has a volume argument showing that contracting this residual forces structure into the visible coordinates, z parallel, which improves global distance fidelity at the cost of local neighborhood fidelity.

Lu: The authors show that tuning this sigma sub perp isn't just about tweaking one parameter; it’s coupled with the grid extent, so you have to tune them together. They also use a Student-t mixture model for the map prior to handle crowding in large datasets.

Meng: I’m looking at those scaling improvements, like when they use a Student-t prior for larger scales on EGFR or hERG datasets. That suggests it’s actually robust enough for real scale problems where crowding is severe.

Lalam: The paper also shows that the exact likelihood from the flow yields an out-of-distribution score that leads the k-NN distance baseline on every chemical dataset, which is a pretty strong signal for new points >

Conclusion: Tom: So to wrap up, NFMap gives us an invertible method that projects new points onto a trained map and carries any map position back to data space. It uses the fixed grid as a common reference canvas without any measurable cost in embedding quality.

Jane: The big implication is that because it’s fully invertible, you aren't limited to just looking at what’s on the map; you can reconstruct the original data exactly from where you are looking, which is powerful for chemical structure analysis.

Lu: It’s easy to extend this framework with a Student-t prior and hierarchical chemical kernel improvements when dealing with larger datasets, which shows the design is extensible.

Meng: From an engineering standpoint, the fact that it's fully invertible is a huge plus because we know we can always trace back any point projected onto the map without losing data >

Lalam: So to summarize NFMap, they combine GTM’s fixed grid with a flow’s invertibility and exact likelihood, providing an open-source package available at https://github.com/hagaxeight/nfmap >

More episodes

← Home