Research map / Time series / Foundation models for time series
Deep Declarative Dynamic Time Warping for End-to-End Learning of Alignment Paths
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train features with TimeErr loss through the optimal continuous GDTW path | DecDTW audio feature learning · Audio-to-score alignment, CQT, Chroma, and Melspec | TimeErr (ms): 19 -> 17; 24 -> 19; 56 -> 16 | Helped |
| Use continuous-time GDTW rather than discrete DTW for alignment | Base features · Audio-to-score alignment | TimeErr (ms): 35 -> 19; 50 -> 24; 122 -> 56 | Helped |
| Train features using Soft-DTW discrepancy without ground-truth path loss | Audio feature learning · Audio-to-score alignment | TimeErr (ms): 35 -> 49; 50 -> 59 | Hurt |
| Train features using feature matching along the ground-truth path without the predicted alignment | Audio feature learning · Audio-to-score alignment | TimeErr (ms): 35 -> 49; 50 -> 59; 122 -> 52 | Hurt |
| Use DILATE path information during feature training | Audio feature learning · Audio-to-score alignment | TimeErr (ms): 35 -> 29; 50 -> 28; 122 -> 26 | Helped |
| Train using DecDTW features but evaluate alignments with discrete DTW instead of GDTW | DecDTW audio feature learning · Audio-to-score alignment, CQT | TimeErr (ms): 17 -> 100 | Hurt |
| Train using DecDTW features but evaluate alignments with discrete DTW instead of GDTW | DecDTW audio feature learning · Audio-to-score alignment, Chroma | TimeErr (ms): 19 -> 45 | Hurt |
| Train sequence features with DecDTW task loss for maximum GPS error | VPR sequence fine-tuning · VPR, maximum error, all conditions, @2m | Test accuracy: Overcast: 10.2 -> 25.4; Sunny: 9.0 -> 22.8; Snow: 6.2 -> 13.5 | Helped |
| Fine-tune sequence features with DILATE path information | VPR sequence fine-tuning · VPR, maximum error, Overcast, @5m | Test accuracy: 44.3 -> 54.5 | Helped |
| Fine-tune with OTAM or ground-truth-path feature cost instead of using base features | VPR sequence fine-tuning · VPR, maximum error | Test accuracy | No clear effect |
| Fine-tune for sequence alignment, then use embeddings for independent image retrieval | VPR DecDTW and DILATE · VPR independent retrieval, maximum error | Test accuracy: improves at ≤ 5m; reduces at > 5m | Mixed |
| Remove warp-refinement iterations to trade solver accuracy for speed (ablation) | GDTW solver · Inference scalability | Inference speed: 4-12x slowdown; mean accuracy loss of only 0.5% | Helped |
Closest papers by meaning
- Perturbed Dynamic Time Warping: A Probabilistic Framework and Generalized Variants (ICLR 2026 · similarity 0.90)
- Deep Canonical Time Warping (CVPR 2016 · similarity 0.90)
- Soft-DTW: a Differentiable Loss Function for Time-Series (ICML 2017 · similarity 0.90)
- Diffeomorphic Temporal Alignment Nets (NeurIPS 2019 · similarity 0.89)
- TimePoint: Accelerated Time Series Alignment via Self-Supervised Keypoint and Descriptor Learning (ICML 2025 · similarity 0.88)
- Improved Learning via k-DTW: A Novel Dissimilarity Measure for Curves (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.