SPLICE: Latent Diffusion over JEPA Embeddings for Conformal Time-Series Inpainting

2026-05-04 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

SPLICE (Self-supervised Predictive Latent Inpainting with Conformal Envelopes) is a modular framework designed for time-series imputation, specifically for power systems, that provides both high reconstruction accuracy and finite-sample reliability guarantees. It integrates a JEPA encoder to map daily load segments into a 64-dimensional latent space, a conditional latent bridge with four sampling modes for generating gap trajectories, an hourly-conditioned decoder to map back to signal space, and Adaptive Conformal Inference (ACI) for coverage-guaranteed prediction bands. The flow-matching variant of SPLICE achieves comparable quality to DDIM with a 5-10x speedup, completing in 5-10 ODE steps. Evaluated on thirteen load datasets (nine proprietary, three UCI Electricity, ETTh1), SPLICE achieved the lowest mean Load-only MSE (0.056), winning 9/12 non-degenerate datasets for 91-day gaps and 18/32 across all gap lengths against five baselines. ACI delivered 93-95% empirical coverage, correcting under-coverage failures of up to 7.5 percentage points observed with static conformal prediction. A pooled JEPA encoder trained on nine feeds demonstrated transferability to four unseen domains, matching or exceeding per-dataset oracles with minimal fine-tuning.

Key takeaway

For AI Scientists and Machine Learning Engineers working on critical time-series imputation tasks, SPLICE offers a robust solution that combines high accuracy with provable uncertainty quantification. You should consider adopting its modular JEPA-based architecture and Adaptive Conformal Inference to ensure reliable predictions and calibrated confidence intervals, especially in non-stationary environments like power systems. This approach can significantly reduce risks associated with over-confident imputations and improve operational planning.

Key insights

SPLICE offers reliable time-series imputation with uncertainty quantification for critical infrastructure like power grids.

Principles

• Latent dynamics models can "imagine" plausible trajectories.
• Adaptive Conformal Inference maintains coverage under distribution shift.
• Modular architectures enable component replacement and transfer learning.

Method

SPLICE uses a JEPA encoder for latent representation, a conditional latent bridge for trajectory generation, an hourly-conditioned decoder for signal reconstruction, and Adaptive Conformal Inference for robust prediction intervals.

In practice

• Use flow-matching for 5-10x faster inference than DDIM.
• Apply ACI to correct under-coverage in non-stationary data.
• Fine-tune only the bridge and decoder for cross-domain transfer.

Topics

• Time-Series Inpainting
• JEPA Embeddings
• Latent Diffusion Models
• Adaptive Conformal Inference
• Power Systems

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• DiffMath: Symbol- and Graph-Aware Latent Diffusion Transformer for Handwritten Mathematical Expression Generation — DiffMath generates HMEG without explicit spatial supervision by using LaTeX structure and a structured latent diffusion model.
• Monthly Diffusion v0.9: A Latent Diffusion Model for the First AI-MIP — A latent diffusion model emulates monthly atmospheric variability with low computational cost for long-timescale climate simulations.
• Uncertainty Quantification for Large Language Diffusion Models — Lightweight, zero-shot uncertainty quantification for LLDMs can achieve high performance with significantly reduced computational cost.
• Variable-Length Tokenization via Learnable Global Merging for Diffusion Transformers — Merging similar tokens enables variable-length tokenization while preserving cross-length representation alignment in Diffusion Transformers.
• Latent Generative Solvers for Generalizable Long-Term Physics Simulation — LGS uses latent space, generative modeling, and an uncertainty knob for stable, generalizable, and efficient long-term PDE simulation.
• Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine — The score function's singularity at low noise drives a two-stage "collapse-and-refine" mechanism for efficient manifold learning in diffusion models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.