Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PLATONT: Learning a Platonic Representation for Unified Network Tomography

Chengze Du, Heng Xu, Zhiwei Yu, Bo Liu, Jialong Li

TL;DR

PLATONT introduces the Platonic Representation Hypothesis to unify network tomography across link estimation, OD traffic, and topology inference by embedding multiple indicators into a shared latent state. It uses multi-indicator alignment via a contrastive learning objective that exactly represents the PMI kernel, along with denoising reconstruction and optional task supervision, to yield compact, structured representations. The approach demonstrates consistent improvements over strong baselines on synthetic and real-world network datasets, achieving higher accuracy and robustness under varying network conditions. Theoretical guarantees on representation alignment and gradient behavior underpin practical gains, suggesting significant potential for cross-task generalization in operational networks.

Abstract

Network tomography aims to infer hidden network states, such as link performance, traffic load, and topology, from external observations. Most existing methods solve these problems separately and depend on limited task-specific signals, which limits generalization and interpretability. We present PLATONT, a unified framework that models different network indicators (e.g., delay, loss, bandwidth) as projections of a shared latent network state. Guided by the Platonic Representation Hypothesis, PLATONT learns this latent state through multimodal alignment and contrastive learning. By training multiple tomography tasks within a shared latent space, it builds compact and structured representations that improve cross-task generalization. Experiments on synthetic and real-world datasets show that PLATONT consistently outperforms existing methods in link estimation, topology inference, and traffic prediction, achieving higher accuracy and stronger robustness under varying network conditions.

PLATONT: Learning a Platonic Representation for Unified Network Tomography

TL;DR

PLATONT introduces the Platonic Representation Hypothesis to unify network tomography across link estimation, OD traffic, and topology inference by embedding multiple indicators into a shared latent state. It uses multi-indicator alignment via a contrastive learning objective that exactly represents the PMI kernel, along with denoising reconstruction and optional task supervision, to yield compact, structured representations. The approach demonstrates consistent improvements over strong baselines on synthetic and real-world network datasets, achieving higher accuracy and robustness under varying network conditions. Theoretical guarantees on representation alignment and gradient behavior underpin practical gains, suggesting significant potential for cross-task generalization in operational networks.

Abstract

Network tomography aims to infer hidden network states, such as link performance, traffic load, and topology, from external observations. Most existing methods solve these problems separately and depend on limited task-specific signals, which limits generalization and interpretability. We present PLATONT, a unified framework that models different network indicators (e.g., delay, loss, bandwidth) as projections of a shared latent network state. Guided by the Platonic Representation Hypothesis, PLATONT learns this latent state through multimodal alignment and contrastive learning. By training multiple tomography tasks within a shared latent space, it builds compact and structured representations that improve cross-task generalization. Experiments on synthetic and real-world datasets show that PLATONT consistently outperforms existing methods in link estimation, topology inference, and traffic prediction, achieving higher accuracy and stronger robustness under varying network conditions.

Paper Structure

This paper contains 23 sections, 2 theorems, 39 equations, 8 figures, 3 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. Related Works
  3. Backgrounds
  4. Network Model and Tomography
  5. Contrastive Learning
  6. Methodology
  7. Problem Formulation
  8. Foundational Concepts
  9. Unified Representation Objective
  10. PlatoNT Design
  11. Design Intuition
  12. Architecture Overview
  13. Learning Objective
  14. Experiment
  15. Experiment Settings
  16. ...and 8 more sections

Key Result

Theorem 1

There exists a feature map $f_X: \mathcal{X} \to \mathbb{R}^d$ such that where $K_{\text{PMI}}$ is the PMI kernel defined in Eq. (eq: pmi_kernel). The shift constant $C$ admits an explicit bound: where $\alpha = \max_{i\neq j}|K_{ij}|$ captures the maximum absolute value of off-diagonal PMI entries, and $K_{ii}$ denotes the diagonal PMI values. Under smoothness assumptions on the indicator co-oc

Figures (8)

  • Figure 1: Illustration of Platonic Network tomography. Latency, congestion, and loss views represent distinct yet complementary projections of a shared latent network state.
  • Figure 2: PlatoNT design intuition. (a) Multi-indicator alignment: Network indicators converge to a shared latent representation through encoder $\phi$. (b) Denoising & tasks regularization: The decoder $\psi$ reconstructs clean indicators from the latent space, then used for task-specific algorithms.
  • Figure 3: Design of PlatoNT. Multiple network indicators are encoded into a shared latent space, aligned via $\mathcal{L}_{\text{align}}$. The latent representation is reconstructed by indicator decoders and the denoised indicators are then used by task-specific algorithms for downstream tomography tasks.
  • Figure 4: Overview of dataset structure.
  • Figure 5: Bias distribution and absolute error comparison across network indicators under noisy environments. Top row: probability density distributions of prediction bias for (a) delay, (b) loss, and (c) bandwidth, each employing dual x-axes to depict values on both linear and logarithmic scales; dashed lines indicate mean values. Bottom row: box plots of absolute errors across different network links over multiple runs. Results are averaged over different noise conditions.
  • ...and 3 more figures

Theorems & Definitions (4)

  • Theorem 1: Exact PMI Kernel Representation
  • Proposition 1: Gradient Reduction via Shared Subspace
  • Proof 1
  • Proof 2