Neural ODE is a family of neural network models that parameterize continuous-time dynamics using ODEs instead of discrete layers — enabling memory-efficient models, continuous normalizing flows, and modeling of irregular time series.

The Core Idea

• Standard ResNet: $h_{t+1} = h_t + f_\theta(h_t)$ (discrete steps)
• Neural ODE: $\frac{dh(t)}{dt} = f_\theta(h(t), t)$ (continuous dynamics)
• Forward pass: Solve the ODE from $t_0$ to $t_1$ using an ODE solver (e.g., Runge-Kutta).
• Backward pass: Adjoint sensitivity method — avoid storing all intermediate states.

Why Neural ODEs Matter

• Memory Efficiency: O(1) memory with adjoint method (vs. O(depth) for ResNets).
• Irregular Time Series: ODE solver naturally handles data sampled at irregular times — no need for fixed step sizes.
• Continuous Normalizing Flows (CNF): Exact density estimation for generative models.
• Adaptive Depth: ODE solver adapts the number of steps based on required accuracy.

Limitations

• Slower than discrete networks — ODE solver requires multiple function evaluations per pass.
• Training is trickier — ODE solver tolerances affect gradients.
• Less expressive than unconstrained ResNets for some tasks.

Connection to Flow Matching

• Flow Matching (2022) extends Neural ODEs for fast, stable generative modeling.
• Used in: Meta's Voicebox (audio), Stable Diffusion 3 (images), AlphaFold 3 (proteins).

Applications

• Time series: Latent ODEs for irregularly sampled clinical data.
• Physics simulation: Modeling physical dynamics with learned ODEs.
• Generative models: Continuous normalizing flows.

Neural ODEs are a theoretically elegant extension of deep learning to continuous dynamics — their influence on Flow Matching makes them relevant to the latest generation of generative models.