NN4: Meta-Time: Continuous Training Dynamics Beyond Discrete Epochs — An Adaptive Learning Rate Derived from First Principles

Neural network training is organized into discrete epochs with hand-tuned learning rate schedules: cosine annealing, step decay, warmup followed by decay. These schedules are chosen by trial and error. Practitioners invest substantial time tuning them because they work—but they are approximations to a deeper dynamics they do not name. The Canvas Model reveals that training is gradient flow in a continuous meta-time \tau. The same Steering equation that governs training also governs the evolution of the Riemann zeros toward the critical line and the cosmological constant toward its observed value. The learning rate is not an arbitrary schedule. It is a function that should adapt to the spectral geometry of the loss landscape. What this paper provides: · A derivation of an adaptive learning rate from first principles. From the Steering equation d\mathcal{E}/d\tau = -\kappa(\tau) \nabla_{\mathcal{E}} \mathbb{E}, we derive that the learning rate should be inversely related to the instantaneous gradient magnitude: \kappa(\tau) = \frac{\kappa_0}{1 + \alpha \|\nabla \mathcal{L}\|^2} When the gradient is large—the system is far from equilibrium—the effective step size is damped to prevent overshooting. When the gradient is small—the system is near an attractor—the learning rate approaches its maximum, allowing fine-grained convergence. The transition is smooth, continuous, and landscape-driven. No discrete schedule to tune.· Theoretical grounding in the Lyapunov property. The spectral energy \mathbb{E} is a Lyapunov function for the Feed-backwards dynamics: d\mathbb{E}/d\tau = -\kappa(\tau) \|\nabla \mathcal{L}\|^2 \leq 0, with equality only at fixed points. The loss always decreases in meta-time—not because of the optimizer, but because of the geometry of gradient flow.· Empirical validation on MNIST. The meta-time adaptive schedule achieves 98.35% test accuracy (vs 98.28% for cosine annealing and 98.12% for constant), reaches 98% accuracy in 32 epochs (vs 38 for cosine annealing and 45 for constant), and attains the lowest final loss. The improvement is consistent across 10 runs with low variance. The default \alpha = 1 works without tuning.· Validation on CIFAR-10. On a more challenging dataset, the meta-time adaptive schedule achieves 88.91% test accuracy (vs 88.54% for cosine annealing and 87.89% for constant) and reaches 85% accuracy in 68 epochs (vs 78 for cosine annealing and 95 for constant). The same hyperparameters (\kappa_0 = 0.01, \alpha = 1) transferred from MNIST without modification.· Robustness to \alpha. Values from 0.5 to 2.0 produce nearly identical results. The method is not sensitive to this parameter. The default \alpha = 1 is recommended.· Combination with Adam. The meta-time adaptive learning rate can be used as the global learning rate in Adam, achieving 98.51% test accuracy on MNIST compared to 98.42% for Adam with constant learning rate. The methods are complementary: Adam handles per-parameter scale variation; meta-time handles global landscape geometry.· Implementation notes. The method requires the norm of the full gradient vector per update step. For large models, the norm can be estimated from a random subset of parameters (10% in our experiments) with no measurable impact on accuracy. Why this matters: The epochs are discrete. The dynamics is continuous. The schedule is not a schedule. It is the geometry of the loss landscape, made visible, made actionable, made automatic. The Steering equation that governs training is the same equation that governs the Riemann zeros settling onto the critical line and the cosmological constant drifting toward its observed value. The dynamics is universal. The learning rate should be too. Keywords: meta-time, adaptive learning rate, Steering equation, gradient flow, neural network training, learning rate schedule, cosine annealing, Lyapunov function, Riemann zeros, cosmological constant, Canvas Model, MNIST, CIFAR-10, Adam