KC-Maps
A deep learning model based on the Kolmogorov-Chapman equations for learning the transition kernels of stochastic differential equations (SDEs). It does not require long trajectory data and is trained using only marginal samples and known coefficients.
Timeline (last 90 days)
KC-Maps (Kolmogorov-Chapman Maps) is a deep learning method that learns stochastic differential equation (SDE) transition kernels without requiring long trajectory data, using only trajectory-free marginal samples together with known drift and diffusion coefficients.
KC-Maps encodes boundary conditions (the transition kernel degenerates to a Dirac δ function when the target time equals the initial time) directly into the neural network architecture, rather than softly penalizing them through the loss function.
KC-Maps uses the Kolmogorov-Chapman equation to construct the training objective, so that the model automatically satisfies multi-step consistency during training.
In experiments on dynamical integration of peptide molecular systems, KC-Maps outperforms classical molecular dynamics simulation in GPU-hour consumption.
The effectiveness of KC-Maps depends on known drift and diffusion coefficients; for real systems with unknown coefficients, parameter estimation is still required first.
KC-Maps has so far been experimentally validated only on peptide molecular systems, and has not yet been validated on more complex macromolecules or non-equilibrium processes.
KC-Maps shifts the deep learning paradigm for SDE simulation from trajectory supervision to distribution matching, reducing the requirements on training data.
All Facts (7)
KC-Maps (Kolmogorov-Chapman Maps) is a deep learning method that learns stochastic differential equation (SDE) transition kernels without requiring long trajectory data, using only trajectory-free marginal samples together with known drift and diffusion coefficients.
50%UnverifiedKC-Maps encodes boundary conditions (the transition kernel degenerates to a Dirac δ function when the target time equals the initial time) directly into the neural network architecture, rather than softly penalizing them through the loss function.
50%UnverifiedKC-Maps uses the Kolmogorov-Chapman equation to construct the training objective, so that the model automatically satisfies multi-step consistency during training.
50%UnverifiedIn experiments on dynamical integration of peptide molecular systems, KC-Maps outperforms classical molecular dynamics simulation in GPU-hour consumption.
50%UnverifiedThe effectiveness of KC-Maps depends on known drift and diffusion coefficients; for real systems with unknown coefficients, parameter estimation is still required first.
50%UnverifiedKC-Maps has so far been experimentally validated only on peptide molecular systems, and has not yet been validated on more complex macromolecules or non-equilibrium processes.
50%UnverifiedKC-Maps shifts the deep learning paradigm for SDE simulation from trajectory supervision to distribution matching, reducing the requirements on training data.
50%