# Soft-DTW: a Differentiable Loss Function for Time-Series

arXiv: https://arxiv.org/abs/1703.01541 · HTML: https://mlsanity.com/paper/1703.01541/

## Key facts

- Cited by 271 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 5 Mar 2017.

## Built on by (most cited first, 30 of 271)

- [NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality](https://mlsanity.com/paper/2205.04421/) (2022)
- [Temporal Alignment Networks for Long-term Video](https://mlsanity.com/paper/2204.02968/) (2022)
- [Tutorial on amortized optimization](https://mlsanity.com/paper/2202.00665/) (2022)
- [Teach a Robot to FISH: Versatile Imitation from One Minute of Demonstrations](https://mlsanity.com/paper/2303.01497/) (2023)
- [Domain Adaptation for Time Series Forecasting via Attention Sharing](https://mlsanity.com/paper/2102.06828/) (2021)
- [NeMF: Neural Motion Fields for Kinematic Animation](https://mlsanity.com/paper/2206.03287/) (2022)
- [Watch and Match: Supercharging Imitation with Regularized Optimal Transport](https://mlsanity.com/paper/2206.15469/) (2022)
- [Deep Differentiable Logic Gate Networks](https://mlsanity.com/paper/2210.08277/) (2022)
- [Warpformer: A Multi-scale Modeling Approach for Irregular Clinical Time Series](https://mlsanity.com/paper/2306.09368/) (2023)
- [Learning by Aligning Videos in Time](https://mlsanity.com/paper/2103.17260/) (2021)
- [Interpretable and Explainable Logical Policies via Neurally Guided Symbolic Abstraction](https://mlsanity.com/paper/2306.01439/) (2023)
- [DBLoss: Decomposition-based Loss Function for Time Series Forecasting](https://mlsanity.com/paper/2510.23672/) (2025)
- [Representation Learning via Global Temporal Alignment and Cycle-Consistency](https://mlsanity.com/paper/2105.05217/) (2021)
- [Drop-DTW: Aligning Common Signal Between Sequences While Dropping Outliers](https://mlsanity.com/paper/2108.11996/) (2021)
- [Online Graph Dictionary Learning](https://mlsanity.com/paper/2102.06555/) (2021)
- [StepFormer: Self-supervised Step Discovery and Localization in Instructional Videos](https://mlsanity.com/paper/2304.13265/) (2023)
- [Patch-wise Structural Loss for Time Series Forecasting](https://mlsanity.com/paper/2503.00877/) (2025)
- [Multi-granularity Correspondence Learning from Long-term Noisy Videos](https://mlsanity.com/paper/2401.16702/) (2024)
- [Fine-grained Temporal Contrastive Learning for Weakly-supervised Temporal Action Localization](https://mlsanity.com/paper/2203.16800/) (2022)
- [Trajectory-Aware Body Interaction Transformer for Multi-Person Pose Forecasting](https://mlsanity.com/paper/2303.05095/) (2023)
- [Pre-training for Speech Translation: CTC Meets Optimal Transport](https://mlsanity.com/paper/2301.11716/) (2023)
- [Video-Text Representation Learning via Differentiable Weak Temporal Alignment](https://mlsanity.com/paper/2203.16784/) (2022)
- [Learning Discriminative Prototypes with Dynamic Time Warping](https://mlsanity.com/paper/2103.09458/) (2021)
- [ScanGAN360: A Generative Model of Realistic Scanpaths for 360circ Images](https://mlsanity.com/paper/2103.13922/) (2021)
- [Inductive and Transductive Few-Shot Video Classification via Appearance and Temporal Alignments](https://mlsanity.com/paper/2207.10785/) (2022)
- [Learnable Expansion of Graph Operators for Multi-Modal Feature Fusion](https://mlsanity.com/paper/2410.01506/) (2024)
- [Temporal-Viewpoint Transportation Plan for Skeletal Few-shot Action Recognition](https://mlsanity.com/paper/2210.16820/) (2022)
- [Time-o1: Time-Series Forecasting Needs Transformed Label Alignment](https://mlsanity.com/paper/2505.17847/) (2025)
- [Learning with Algorithmic Supervision via Continuous Relaxations](https://mlsanity.com/paper/2110.05651/) (2021)
- [Uncertainty-DTW for Time Series and Sequences](https://mlsanity.com/paper/2211.00005/) (2022)

## Abstract (arXiv)

We propose in this paper a differentiable learning loss between time series, building upon the celebrated dynamic time warping (DTW) discrepancy. Unlike the Euclidean distance, DTW can compare time series of variable size and is robust to shifts or dilatations across the time dimension. To compute DTW, one typically solves a minimal-cost alignment problem between two time series using dynamic programming. Our work takes advantage of a smoothed formulation of DTW, called soft-DTW, that computes the soft-minimum of all alignment costs. We show in this paper that soft-DTW is a differentiable loss function, and that both its value and gradient can be computed with quadratic time/space complexity (DTW has quadratic time but linear space complexity). We show that this regularization is particularly well suited to average and cluster time series under the DTW geometry, a task for which our proposal significantly outperforms existing baselines. Next, we propose to tune the parameters of a machine that outputs time series by minimizing its fit with ground-truth labels in a soft-DTW sense.