Neural Networks as Controlled Dynamical Systems

This direction is split into other directions: mean-field optimal control. Its last version stays readable and citable here.

Version 1 of 136 papers41 conceptsfrozen 2026-10-04

For decades, training a deep neural network has been viewed primarily through the lens of optimization: we define a loss landscape and use gradients to "roll" down to the lowest point. While successful, this perspective often treats the internal mechanics of the network as a black box and the training process as a series of heuristic updates.

A growing body of research suggests a more rigorous alternative: viewing neural networks as dynamical systems and their training as an optimal control problem. In this framework, the forward pass is a trajectory through space, and the weights are not just parameters, but "controllers" steering that trajectory toward a desired destination.

The Forward Pass as a Trajectory

The first step in this shift is reimagining the architecture itself. In a standard feedforward network, we pass data through a sequence of layers. If we view each layer as a discrete time step, the transformation of data becomes a dynamical system. As Yiping Lu and colleagues noted in Beyond Finite Layer Neural Networks, deep architectures like ResNets can be seen as Euler discretizations of continuous differential equations.

When we treat the forward pass as a continuous flow, we gain access to the tools of control theory. For instance, Qianxiao Li and Zhuotong Chen introduced Close-loop Control Neural Networks (CLC-NN), which add a feedback loop to the forward pass Towards Robust Neural Networks via Close-loop Control. Just as a thermostat adjusts a heater to maintain a temperature, this feedback loop keeps internal activations on a "safe" manifold, making the network naturally resilient to adversarial perturbations.

Training as Steering

If the forward pass is a trajectory, then training is the act of finding the best way to steer it. This is the classic Optimal Control Problem (OCP). In the late 80s, Oluseyi Farotimi, Amir Dembo, and Thomas Kailath first formalized this, defining a performance index to derive optimal weight trajectories Neural Network Weight Matrix Synthesis Using Optimal Control Techniques.

In this view:

  • States are the hidden activations of the network.
  • Controls are the weights and biases of each layer.
  • The Objective is to steer the initial state (input data) to a target state (labels) while minimizing "effort" (regularization).

This framing allows researchers to move beyond simple gradient descent. For example, Guan-Horng Liu and Evangelos Theodorou have developed optimizers like OCNOpt and DDPNOpt that use Differential Dynamic Programming Differential Dynamic Programming Neural Optimizer. Instead of a global gradient update, these methods compute layer-wise "feedback policies," which can lead to faster convergence and better stability in complex architectures like Neural ODEs.

Pontryagin’s Principle and Discrete Weights

One of the most powerful tools in control theory is Pontryagin’s Maximum Principle (PMP). PMP provides necessary conditions for an optimal control trajectory. In deep learning, this offers a way to train networks where gradients are unavailable or poorly behaved.

Qianxiao Li and Shuji Hao applied PMP to develop the Method of Successive Approximations (MSA) An Optimal Control Approach to Deep Learning and Applications to Discrete-Weight Neural Networks. Because MSA doesn't rely on backpropagating gradients through the entire chain, it is uniquely suited for training networks with binary or ternary weights—architectures where standard SGD typically fails.

Stability and Convergence Guarantees

Control theory also provides a language for stability. Standard training often suffers from "exploding" or "vanishing" gradients, which are essentially instabilities in the system's dynamics. By applying Lyapunov Stability Analysis, researchers like Dahao Tang have created dynamic learning rate controllers (SGD-DLR) that mathematically guarantee the optimization process will remain stable and converge smoothly Stability-Driven CNN Training with Lyapunov-Based Dynamic Learning Rate.

Similarly, the "burn-in" phase in Recurrent Neural Networks (RNNs) can be explained by the turnpike phenomenon in optimal control. Julian D. Schiller and colleagues showed that optimal trajectories tend to stay near a steady state for most of their duration, providing a theoretical justification for why ignoring early sequence outputs improves performance Tuning the burn-in phase in training recurrent neural networks improves their performance.

A Unified Physical Framework

Perhaps the most ambitious aspect of this research is the unification of learning rules. Alessandro Betti and Marco Gori have proposed Hamiltonian Learning (HL), which treats the evolution of weights as a physical process governed by Hamiltonian mechanics A Unified Framework for Neural Computation and Learning Over Time.

In this framework, standard Backpropagation is not a standalone algorithm, but a specific, constrained instance of a more general continuous-time dynamical system. By viewing learning as a physical trajectory that minimizes "kinetic energy" (weight changes) alongside error, we can derive new, spatiotemporally local learning rules that are more biologically plausible and efficient for online learning from data streams.

By treating neural networks as controlled dynamical systems, we move from "tuning" a black box to "engineering" a predictable, stable, and robust flow of information.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_networks_as_controlled_dynamical__12770490,
  title = {Neural Networks as Controlled Dynamical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-networks-as-controlled-dynamical-systems/rdr_12770490920d4fbeb71d72bdeb28b74d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_neural_networks_as_controlled_dynamical__12770490_v1,
  title = {Neural Networks as Controlled Dynamical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-networks-as-controlled-dynamical-systems/rdr_12770490920d4fbeb71d72bdeb28b74d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}