Tension Adaptation Versus Scheduler Effects on CIFAR-10: A Matched Ablation Study
License: CC BY 4.0 License
DOI: 10.21203/rs.3.rs-9089751/v1
Keywords: Artificial Intelligence and Machine Learning, CIFAR-10, uncertainty quantification, diagnostic ablation, per-sample trajectory analysis, epistemic calibration, learning-rate scheduling, adaptive optimization, confidence-error decomposition
Abstract
Adaptive training schemes often combine several control components at once, making it unclear which mechanism is actually responsible for improved learning. This paper isolates that question on CIFAR-10 with a matched 4-condition ablation over a fixed convolutional classifier and fixed training protocol. We vary only two components: a scheduler/warmup controller and a per-batch tension-adaptation mechanism that adjusts optimization behavior from loss-tension signals between successive updates. Under the refreshed checkpoint-backed run set, the base condition reaches 68.89% final accuracy, scheduler-only improves to 71.46%, tension-only reaches 74.42%, and the combined condition reaches 74.66%. Per-sample trajectory populations show the same pattern: relative to the base run, the stronger conditions reduce persistent uncertainty and increase uncertainty-to-correct conversion, but they do so in different ways. The combined condition preserves the lowest persistent confident-error mass, while tension-only carries the largest uncertainty-to-correct conversion count. A diagnostic exclusion sweep over CVS tail populations further shows that weak-specification cases explain only a modest part of the remaining ceiling, whereas the volatile cluster is load-bearing for learning rather than disposable noise. In this setting, scheduler effects are real but secondary, tension adaptation remains the dominant mechanism, and certainty-validity diagnostics are most informative when read together with coverage and per-sample trajectory populations.