Neural ordinary differential equation (neural ODE)

A neural ODE parameterizes the instantaneous rate of change of a latent state by a neural network, dx/dt = f_theta(x, t, u), and obtains the trajectory by numerical integration; gradients are computed either by backpropagating through the solver or via the adjoint sensitivity method, which integrates a second ODE backward in time with memory independent of trajectory length. Placed inside a sequential VAE, the neural ODE becomes a continuous-time dynamics prior, giving a smooth, arbitrarily nonlinear vector field over the latent space in place of a discrete-time RNN recurrence.

The continuous-time formulation is a natural fit for neural population dynamics: it handles irregular or missing samples gracefully, ties directly to the flow-field and fixed-point language physiologists already use, and lets one read the learned f_theta as an estimated vector field. Practical costs are real — stiff dynamics force many solver steps, adjoint accuracy can be delicate, and, as with any flexible dynamics model, fitting the data well does not by itself recover the true governing equations. Variants such as ODE-RNNs and latent-SDE models add jumps at observation times or stochasticity to the drift.