The Latent Information Geometry of Jet Classification

This paper introduces a framework based on differential and information geometry to analyze the latent representations of machine learning models, applying concepts like curvature and nonmetricity to interpret the physics underlying quark-gluon and fat jet classification tasks.

The Gist

Imagine you have a super-smart robot that looks at pictures of particles smashing together in a giant collider (like the Large Hadron Collider) and has to guess: "Is this a quark jet or a gluon jet?"

The robot is incredibly good at this. It gets the answer right almost every time. But here's the problem: We don't know how it knows. It's a "black box." It sees patterns we can't see, and if we can't explain its logic, we can't fully trust it when it's making decisions about new physics.

This paper introduces a new way to peek inside the robot's brain. Instead of just looking at the answer, the authors use a branch of math called Information Geometry to map out the "shape" of the robot's thinking process.

Here is the breakdown using simple analogies:

1. The Robot's "Dream World" (Latent Space)

When the robot processes a particle collision, it doesn't keep the raw data. It compresses it into a tiny, simplified summary. Think of this like a dream.

• You see a chaotic scene in real life, but in your dream, it's a simple, abstract feeling or image.
• The robot creates a "latent space" (a dream world) where every type of particle jet lives at a specific coordinate.
• The Goal: We want to understand the landscape of this dream world. Are the "Quark" mountains far away from the "Gluon" valleys? Is the path between them a straight road or a winding cliff?

2. The Mapmaker's Toolkit (Information Geometry)

Usually, we think of space as flat (like a sheet of paper). But the authors argue that the robot's dream world is curved and warped, like a crumpled piece of paper or a saddle.

To understand this shape, they use three main tools:

• The Stretchy Ruler (Fisher Metric): Imagine a ruler that changes length depending on where you are. In some parts of the robot's brain, a tiny step means a huge change in meaning. In other parts, you can walk miles and nothing changes. This "stretchy ruler" tells us where the robot is most sensitive to changes.
• The Twisted Compass (Curvature): If you walk in a perfect circle in this dream world, do you end up facing the same direction? In a flat world, yes. In the robot's world, maybe you end up facing the opposite way. This "twist" tells us how complex the relationships between particles are.
• The Stretchy Fabric (Non-metricity): This is the paper's big discovery. Imagine a fabric that not only bends but also stretches or shrinks as you walk across it. The robot's brain uses this "stretching" to encode information. It's like the robot says, "To turn a Gluon into a Quark, you don't just walk; you have to stretch the fabric of reality."

3. The "Decision Line" (The Boundary)

The most important part of the robot's job is drawing a line between "Quark" and "Gluon."

• The Old Way: We just looked at the line and said, "Okay, it's here."
• The New Way: The authors show that the robot doesn't just draw a line; it builds a fortress around it.
  • Right at the line, the "stretchy ruler" goes crazy (it stretches infinitely). This means the robot is hyper-sensitive here; a tiny nudge changes the answer completely.
  • The "stretchy fabric" (non-metricity) is strongest here. The robot is using complex geometry to make sure it doesn't make mistakes.

4. The "Magic Scalars" (The New Tools)

The authors invented four new numbers (scalars) to measure this weird geometry. Think of them as thermometers for the robot's confusion.

• Scalar 1, 2, and 3: These measure how "lopsided" or "skewed" the robot's thinking is. If the number is high, the robot is using a very complex, non-standard way to tell the difference.
• The Discovery: They found that the robot doesn't use "curvature" (bending) to solve the problem. Instead, it uses non-metricity (stretching/shrinking). It's like the robot solves the puzzle by stretching the rubber sheet of space rather than bending it.

5. Real-World Application: The Particle Zoo

They tested this on three scenarios:

1. Digits 1 vs. 7: A simple test. The robot learned that the "angle" of the line was the most important thing to stretch the space around.
2. Quarks vs. Gluons: This is the hard physics problem. They found that the robot learned to separate them based on multiplicity (how many tiny particles are in the jet). The "stretchy ruler" showed that the robot is most sensitive to the number of particles right at the decision line.
3. Top vs. Z vs. Quark/Gluon: A three-way fight. They found that the robot treats the "Top" particle as very far away from the others (easy to distinguish), but the "Z" and "Quark/Gluon" are close neighbors. The geometry showed that to turn a "Top" into a "Quark," the robot has to pass through a "Z-like" state first, like a stepping stone.

The Big Picture

This paper is like giving us X-ray glasses for AI.

Instead of just trusting the robot because it gets high scores, we can now look at the shape of its brain. We can see:

• Where it is confident (flat, calm areas).
• Where it is confused (highly stretched, twisted areas).
• What physical features it actually cares about (the direction of the stretch).

Why does this matter?
If we understand the "shape" of the robot's logic, we can fix it if it's wrong, trust it when it's right, and even use it to discover new laws of physics that we haven't thought of yet. It turns the "black box" into a transparent, navigable map.

Technical Summary

Here is a detailed technical summary of the paper "The Latent Information Geometry of Jet Classification" by Kuntz et al.

1. Problem Statement

Modern machine learning (ML) in high-energy physics, particularly for jet tagging at the Large Hadron Collider (LHC), relies heavily on deep neural networks (DNNs) to classify particle jets (e.g., distinguishing quark jets from gluon jets, or identifying top quarks). While these networks achieve high performance, they often function as "black boxes."

• The Core Issue: It is difficult to understand what physical features the network relies on, how the latent representation aligns with known physics observables, and how the network encodes correlations.
• The Gap: Standard Euclidean analysis of latent spaces is insufficient because the networks learn complex, non-linear manifolds. There is a need for a mathematically rigorous framework to analyze the geometry of these learned representations to ensure they are physically sound and explainable.

2. Methodology: Information Geometry

The authors propose applying Information Geometry to analyze the latent spaces of Variational Autoencoders (VAEs) equipped with classifier heads. They treat the network's output probabilities as a parametric family of distributions forming a statistical manifold.

Key Theoretical Concepts:

• Fisher Information Metric ($g_{ij}$): Used as the Riemannian metric to measure distances between probability distributions in the latent space. It quantifies the sensitivity of the classifier output to changes in latent variables.
• Dual Connections ($\nabla^{(\pm 1)}$): Instead of the standard Levi-Civita connection, the authors utilize the dual connections derived from the Kullback-Leibler (KL) divergence. These connections are torsion-free but not metric-compatible.
• Nonmetricity and the Amari-Chentsov Tensor (ACT): The failure of the dual connections to preserve the metric is quantified by the nonmetricity tensor, which in this context is the Amari-Chentsov tensor ($C_{ijk}$). This tensor measures the skewness of the likelihood distribution.
• Geodesics vs. Autoparallels: The paper distinguishes between length-minimizing geodesics (Levi-Civita) and autoparallels (curves where the tangent vector is parallel-transported by the dual connection). The deviation between these paths reveals the non-metric nature of the learned geometry.

Novel Scalars:
To quantify the geometry without relying on coordinate-dependent measures, the authors introduce four new scalars derived from the ACT:

1. $C_1 = C_{ijk}C^{ijk}$: The full contraction of the ACT (analogous to the Fubini-Pick invariant).
2. $C_2 = \tau_i \tau^i$: The contraction of the Chebyshev field (the trace of the ACT).
3. $C_3 = \tilde{C}_{ijk}\tilde{C}^{ijk}$: The contraction of the traceless part of the ACT. Non-zero $C_3$ indicates an irreducible asymmetry (non-Gaussianity) that cannot be removed by conformal transformations.
4. $C_4$: A scalar related to the difference between the Levi-Civita Ricci scalar and the dual Ricci scalars.

3. Key Contributions

1. Framework for Latent Geometry: Established a method to map the latent space of ML classifiers to a statistical manifold using Fisher information and dual connections.
2. New Geometric Diagnostics: Proposed the scalars $C_1, C_2, C_3$ to detect decision boundaries and model complexity. They demonstrated that for binary classifiers, the geometry is effectively one-dimensional and encoded in nonmetricity (ACT) rather than curvature (Ricci scalars vanish).
3. Feature Alignment Analysis: Developed a method to correlate the latent geometry with physical observables by computing the Fisher directional derivative. This identifies which physical feature (e.g., jet mass, multiplicity) drives the classification decision in a specific region of the latent space.
4. Application to Jet Physics: Applied these methods to:
   • A toy MNIST dataset (digits 1 vs. 7) to validate the methodology.
   • Binary Quark-Gluon Tagging: Analyzing the latent structure of quark vs. gluon jets.
   • Three-Class Jet Tagging: Distinguishing between Quark/Gluon ($q/g$), $Z \to q\bar{q}$, and Top ($t \to b q \bar{q}'$) jets.

4. Results

Toy Model (MNIST 1 vs. 7):

• The Fisher Frobenius norm peaks at decision boundaries, indicating high sensitivity.
• The scalars $C_1, C_2, C_3$ sharply trace the decision boundary.
• The "Chebyshev field" flow is orthogonal to the decision boundary, representing the fastest path to change a class label.
• The geometry is dual-flat (Ricci scalars vanish), meaning information is encoded entirely in nonmetricity.

Quark-Gluon Tagging:

• Geometry: The classifier geometry is effectively one-dimensional (elongated ellipses parallel to the decision boundary), while the decoder geometry is two-dimensional.
• Feature Dominance: The dominant feature separating quarks and gluons is constituent multiplicity ($n_{PF}$).
• Geodesic Analysis: Moving from gluon to quark regions along geodesics shows a linear evolution of multiplicity. In specific low-density regions, the classifier relies on radiation spread ($w_{PF}$) or fragmentation patterns ($S_{frag}$).
• Nonmetricity: The scalar $C_4$ vanishes, confirming that the information is encoded in the nonmetricity tensor (skewness) rather than curvature.

Three-Class Tagging ($q/g$ vs. $Z$ vs. $t$):

• Topology: The latent space forms a triangular structure. The distance between Top jets and the other two classes is significantly larger (higher Fisher-Rao distance) than the distance between $Z$ and $q/g$, consistent with the superior performance of top taggers.
• Transition Paths:
  • The transition from $Z$ (2-prong) to Top (3-prong) is dominated by $\tau_{32}$ (N-subjettiness ratio).
  • The transition from Top to $q/g$ (1-prong) involves a complex path where the jet may pass through intermediate $Z$-like features (2-prong) before becoming 1-prong. This suggests a hierarchical evolution of jet substructure: $t \to Z \to q/g$.
• Alignment: High cosine similarity between classifier and decoder metrics along decision boundaries indicates the network learns physically meaningful features (e.g., N-subjettiness ratios and jet mass).

5. Significance

• Explainability: This work moves beyond "feature importance" (like SHAP values) to a geometric understanding of how a network organizes data. It reveals that networks encode decision boundaries via nonmetricity (skewness) rather than curvature.
• Physics Validation: By showing that the latent geometry aligns with known QCD observables (multiplicity, N-subjettiness, Sudakov form factors), the paper validates that deep learning taggers are learning the underlying physics rather than just detector artifacts.
• Simulation Gap: Understanding the latent geometry helps in bridging the gap between simulation-based training and real data. If the geometry is understood, one can better correct for simulation mismodeling.
• Generalizability: The proposed scalars ($C_1, C_2, C_3$) and the information geometry framework are applicable to any classification task where the latent space is learned, offering a new tool for auditing and interpreting neural networks in scientific domains.

In summary, the paper successfully translates concepts from General Relativity and Information Geometry into a practical toolkit for analyzing the "black box" nature of jet classifiers, revealing that the decision-making process is governed by non-metric geometric structures that align with fundamental QCD principles.