Abstract
Transformers consuming multi-channel scalar signals must embed $C$ simultaneous values into one $d_{\text{model}}$-dimensional vector per time step. We empirically audit eight input encoders -- spanning a shared-scalar baseline, per-channel linear projections, an orthogonality regulariser, a nonlinear MLP stem, block-partitioned concatenation, channel-independent and channel-as-token architectures, and a projected positional encoding -- on a synthetic benchmark designed to make channel identity informative and on ETTh1 as a real-data check, measured in next-step negative log-likelihood (NLL). The headline is one of practical near-equivalence within a wide "top tier": the standard per-channel linear projection (this http URL(C, $d_{\text{model}}$)) matches every alternative in that tier up to small, statistically real but practically modest, differences. Two encoders lose decisively: the shared-scalar baseline, which collapses for information-theoretic reasons we make explicit, and the channel-independent PatchTST-spirit baseline, which underperforms on both benchmarks and overfits universally on the synthetic one. Paired tests resolve two small gaps: projecting the sinusoidal positional encoding through a learned linear layer edges the rest at small $C$, with a direct geometric probe showing the mechanism is positional-channel orthogonalisation; a nonlinear MLP stem edges them at the largest $C$ we test, with the gap shrinking under more training data. The practical recommendation is to use this http URL(C, $d_{\text{model}}$) by default and reach for something more elaborate only when the task at hand gives a real reason to do so. Code and data to reproduce every experiment in this paper are available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2606.04752
