BCIwiki (bciwiki.com) — Researchers affiliated with the National University of Singapore, KU Leuven and MIT introduced ATEMPO to identify informative lags between time series. The manuscript was posted to arXiv on October 27, 2026. Across three neural benchmark tasks, selected lags increased mean R² by 4.1% relative to the strongest competing baseline.
The method uses transfer entropy, which measures how much a source signal's history adds to prediction beyond the target signal's own history. ATEMPO is a diffusion-based estimator that samples source and conditioning depths during training and explicitly conditions on temporal position. A single score network can therefore estimate the full transfer entropy surface without refitting; differences across that surface recover when information arrives. Such predictive information relationships do not by themselves establish causality.
On synthetic datasets with ground-truth transfer entropy, ATEMPO achieved the lowest mean relative error among the compared baselines. On datasets with known lags, its lag-detection accuracy was 0.98±0.04. It required a single fit across source depths K=6 and target depths L=3, whereas existing estimators required 18 separate fits for the same K×L grid. Across three systems with time-varying lags, ATEMPO was the only compared estimator to maintain perfect lag-detection accuracy at every evaluated window length.
Across three Neural Latent Benchmark tasks using existing monkey-experiment data, ATEMPO-selected lags achieved the highest or equal-best R², the coefficient of determination used to assess predictive fit. Mean R² improved by 4.1% relative to the strongest competing baseline. This result concerns prediction performance in the evaluated tasks and does not establish improved clinical efficacy for BCIs.