Neural Ordinary Differential Equations (Neural ODEs) are a family of deep learning architectures that model the hidden state dynamics as a continuous-time differential equation — dh/dt = f(h, t; θ) — replacing the discrete layer-by-layer transformation of ResNets with continuous-depth evolution integrated by a numerical ODE solver, enabling adaptive-depth computation, exact invertibility for normalizing flows, memory-efficient training via the adjoint method, and natural modeling of continuous-time processes from irregularly sampled data.

The Continuous Depth Insight

Residual networks compute: h_{l+1} = h_l + f(h_l, θ_l)

This is equivalent to Euler's method for solving an ODE with step size 1. Neural ODEs generalize this to the continuous limit:

dh/dt = f(h(t), t; θ), h(0) = x, output = h(T)

The transformation from input x to output h(T) is the solution to an ODE over the interval [0, T]. The function f (implemented as a neural network) defines the vector field — the "velocity" at each point in state space. The ODE solver (Dopri5, Adams, or Euler) integrates this field.

Key Properties and Capabilities

Adaptive computation depth: The ODE solver adapts its step count based on the dynamics' stiffness. Simple inputs require few solver steps (fast inference); complex inputs requiring precise integration take more steps. This is the first neural architecture where computation automatically scales with input difficulty.

Memory-efficient training via the adjoint method: Standard backpropagation through the ODE solver requires storing O(N) intermediate states where N is the number of solver steps — memory-intensive for deep integration. The adjoint sensitivity method avoids this: it computes gradients by solving a second ODE backward in time, using O(1) memory regardless of integration depth.

The adjoint ODE: da/dt = -a(t)^T · ∂f/∂h, where a(t) = ∂L/∂h(t) is the adjoint state.

Exact invertibility: The ODE defining the forward pass is exactly invertible — given h(T), recover h(0) by integrating backward. This enables Neural ODEs to be used as normalizing flows (exact density computation) without the architectural constraints of coupling layers required by RealNVP or Glow.

Continuous-time input modeling: For sequences with irregular time stamps (medical records, sensor data with gaps), Neural ODEs naturally model state evolution between observations without interpolation or masking.

ODE Solver Options

The choice of solver dramatically affects training stability and speed. Dopri5 is the default for most applications.

Latent Neural ODEs for Time Series

Latent Neural ODEs combine Neural ODEs with the VAE framework for generative modeling of irregularly-sampled time series: 1. Encoder (RNN or attention) maps observations to initial latent state z₀ 2. Neural ODE integrates z₀ forward to prediction times 3. Decoder produces observations from latent state 4. Training: ELBO with reconstruction loss + KL regularization

This enables generation at arbitrary time points, uncertainty quantification, and imputation of missing values — critical capabilities for clinical time series.

Limitations and Challenges

• Training instability: Stiff ODE dynamics produce small maximum step sizes, dramatically increasing training cost and causing gradient issues
• Solver overhead: Even with adjoint method, inference requires multiple function evaluations per ODE step — slower than equivalent discrete networks for standard tasks
• Trajectory crossing: Vector field f must be Lipschitz continuous (guaranteeing unique solutions), which prevents trajectories from crossing — limiting expressiveness for complex transformations (addressed by Augmented Neural ODEs)

Neural ODEs sparked a research program connecting differential equations and deep learning, producing CfC networks (closed-form dynamics), Neural SDEs (stochastic), Neural CDEs (controlled), and continuous normalizing flows — each addressing specific limitations while preserving the core insight that deep learning and dynamical systems theory share fundamental mathematical structure.