Training a neural network is like sledding down a hill: you need the right speed to avoid getting stuck or flying off the track. In the popular momentum method (like inertia), a constant value of 0.9 is typically used. The authors revisited the physics of damped oscillations and proposed changing momentum during training: initially large, then decreasing according to a formula where momentum equals one minus two times the square root of the learning rate.
They tested it on a standard task: a ResNet-18 neural network learning to recognize images from the CIFAR-10 dataset. With the new momentum schedule, the model reaches 90% accuracy 2.34 times faster than with constant 0.9. However, the final accuracy is slightly lower — by 0.46 percentage points. So they applied a hybrid: critical damping until 90%, then switching to constant momentum. This eliminates the accuracy loss and retains the speedup: final accuracy is 95.45%, and it reaches 90% accuracy 2.4 times faster.
In the second part of the paper, the authors tested the idea of selecting layers for fine-tuning based on error gradients. It turned out that gradient-based selection is no better than random. But there is a useful signal: if a layer has the lowest gradient, it's definitely not worth touching — interfering with it can be very harmful. So gradients help avoid bad decisions, not find good ones.
as written by the authors arXiv:2603.28921
The critical damping condition of the damped harmonic oscillator model of SGD with momentum (Qian, 1999) yields a momentum schedule with no tuned hyperparameters: mu(t) = 1 - 2*sqrt(alpha(t)). Across five seeds on ResNet-18/CIFAR-10 (200-epoch cosine schedule) it reaches 90% test accuracy 2.34x faster than constant mu=0.9 (range 1.71-2.86x, 5/5 seeds, one-sided paired t-test p=4e-4), at the cost of a real final-accuracy deficit of 0.46 pp (5/5 seeds, p=0.009). A short-schedule control rules out a schedule-length artifact: compressed baselines either pay 0.5-0.9 pp of accuracy or stay slower to 90% at equal accuracy. A hybrid recipe -- critical-damping momentum until 90%, then constant mu=0.9 -- removes the deficit and keeps the speedup: 95.45 +/- 0.05% final accuracy at 2.4x faster progress to 90% (n=5). The speedup generalizes across architectures (VGG-16 without skip connections: 1.72x, n=3); on CIFAR-100 early gains persist (2-4x to mid-training thresholds) but the accuracy cost grows (-1.7 pp), compressing the accuracy-matched gain to 1.14x. We also report an exhaustive negative result on surgical layer selection. Version 2 of this paper claimed that gradient attribution on misclassified images selects which layers to retrain; running the identical correction protocol on all 35 combinations of 3-of-7 layer groups ranks the selected triple 11th of 35 (exact p=0.31) -- no better than random. What survives is weaker but real: combinations containing the top-ranked layer outperform the rest (+6.2 vs -1.4 mean net error reduction), and the bottom of the gradient-norm ranking reliably predicts the most harmful interventions (down to -20 net errors). Gradient attribution on errors is a harm-avoidance signal, not a selector of repair targets. We release the full 35-combination landscape as a baseline for layer-selection claims.
Momentum in optimizers is named after the physical concept of moment of inertia: it helps overcome local minima, just as a sled glides over small bumps thanks to its accumulated speed.