2026 workshop program
Time encoding for irregularly sampled light curves
The work
Abstract
Self-supervised representation learning is the standard response to the labeling bottleneck on large astronomical surveys, but evaluating what these encoders have actually learned about time — as opposed to class — is harder than it looks. Period regression R^2 on a multi-class catalog conflates two very different behaviors: the encoder identifying the source’s class and the probe predicting the class-mean period, versus the encoder reading the period off an individual source’s photometry. We propose decomposing predicted-log-period variance into between-class (f_cls) and within-class (rho_within) components to separate them. Applied to a range of self-supervised encoders on irregularly sampled ZTF light curves — including a 46M cross-domain pretrained foundation model, a 4.4M cadence-as-channel BiGRU, and a continuous-time SSL transformer with explicit pairwise time-bias attention — the diagnostic returns the same answer everywhere: 60–70% of period R^2 is class-template matching, and rho_within tops out near 0.26. We then survey several time-encoding strategies for closing this gap — per-step Δt channels, random Fourier features, continuous-time attention bias, and period-sensitive contrastive objectives — and find that improvements on classification do not translate to improvements on the within-class period diagnostic. Learning absolute time from irregular astronomical photometry remains an open problem; we offer the diagnostic as a way for the field to track progress on it.