Soft-DTW
A differentiable variant of DTW that replaces the min operation in dynamic programming with a soft-min operator based on log-sum-exp smoothing. A temperature parameter γ controls the degree of smoothing, enabling DTW-based losses to participate in gradient-based optimization for sequence learning tasks.
Timeline (last 90 days)
Soft-DTW replaces the min operation in dynamic programming with a soft-min operator (a smooth approximation based on log-sum-exp), making the metric differentiable.
Soft-DTW controls the degree of smoothing via a temperature parameter γ; when γ→0 it degenerates to the original DTW, and as γ increases the gradients become smoother but the metric accuracy is somewhat reduced.
Neither Soft-DTW nor CDTW has resolved the computational efficiency problem, and the computational overhead remains high on large-scale datasets.
All Facts (3)
Soft-DTW replaces the min operation in dynamic programming with a soft-min operator (a smooth approximation based on log-sum-exp), making the metric differentiable.
50%UnverifiedSoft-DTW controls the degree of smoothing via a temperature parameter γ; when γ→0 it degenerates to the original DTW, and as γ increases the gradients become smoother but the metric accuracy is somewhat reduced.
50%UnverifiedNeither Soft-DTW nor CDTW has resolved the computational efficiency problem, and the computational overhead remains high on large-scale datasets.
50%