Sensor-Conditioned Representation Learning via Scene-Relevant Observation Quotients

2026-06-15

AI Summary

Under sensor-constrained conditions, conventional representation learning struggles to disentangle intrinsic scene variations from nuisance factors or sensor artifacts, often leading to distorted representations. This work formally establishes, for the first time, a correctness criterion for representations under sensor constraints and introduces a decoupling learning approach based on the scene-relevant Observation Quotient (OQ). We propose the OQ-TSAE framework, which integrates the observation quotient as a supervisory signal into a Tucker-structured autoencoder by aligning quotient spaces and normalizing nuisances, thereby explicitly separating scene-related and interference factors. Experiments demonstrate that our method significantly outperforms baseline approaches—including reconstruction-, metric-, and contrastive-learning methods—on both controlled simulations and real radar data, achieving consistent improvements in representation fidelity, robustness, and downstream task performance.

Abstract

Learned representations in intelligent sensing systems are often evaluated by reconstruction fidelity or downstream prediction accuracy, but these criteria do not specify which latent distinctions are justified by the sensing process. In sensor-conditioned environments, nuisance factors can change measurements without changing the scene, while distinct scenes may be indistinguishable under limited sensing capability. This paper formulates sensor-conditioned representation correctness as preserving sensing-supported scene distinctions while suppressing nuisance-induced and sensor-unsupported variation. We introduce the scene-relevant observation quotient, a representation target induced by sensing-supported distinguishability after nuisance canonicalization, and develop Observation-Quotient Tucker-Structured Autoencoding (OQ-TSAE), a scene-nuisance factorized framework with diagnostics for false distinction, false merge, nuisance sensitivity, and latent ordering consistency. Experiments on a controlled benchmark show that quotient-consistent supervision improves representation-correctness diagnostics over reconstruction-oriented, metric-learning, and contrastive-learning baselines. Sensitivity, perturbation, and ablation studies show the importance of quotient-aligned supervision, reliable quotient relations, and quotient geometry. Complementary real-radar experiments show that a reconstruction-only OQ-TSAE variant retains competitive downstream utility, robustness under observation degradation, and low seed-to-seed variability. These results suggest that sensor-conditioned representations should be evaluated not only by predictive utility, but also by whether their latent geometry preserves sensing-justified scene distinctions.

Problem

• sensor-conditioned representation
• scene distinction
• nuisance factors
• representation correctness
• observation quotient

Innovation

• sensor-conditioned representation
• observation quotient
• nuisance canonicalization
• Tucker-structured autoencoding
• representation correctness