arXiv subjects

Gongyue Zhang

Publications and source records attributed to Gongyue Zhang.

5 recordsLinked to original sources

arXiv2026

The Conflict Between Logic and Memory: Training Conditions for Optimizer-Dependent Rule Acquisition

Optimizers can fit the same task while acquiring different generalizing relations. We study the training conditions governing these differences in single-hidden-layer ReLU networks, combining composite evidence tasks, parameter-level interventions, and a three-seed strict-parity scan. Our central finding is that nuisance-connected trainability reshapes both shared failures and relative optimizer advantages. In a nuisance-heavy task, all twenty tested optimizer configurations remain near chance on the hardest stage. Retaining every input but fixing nuisance-connected first-layer weights at initialization raises that stage's accuracy from approximately 50\% to 70.56\%, 68.47\%, and 69.00\% for momentum SGD, Adam, and Muon. Masking the same inputs only after full training does not recover this performance. On a separate pairwise-mode task, background freezing reduces Muon's rare-mode advantage over momentum SGD by 12.48 percentage points, while the target and mode frequencies remain fixed. Each intervention is evaluated under a common validation-selection protocol with condition-specific learning rates and checkpoints. A strict-parity sweep over orders 1--20 provides a complementary reference without spurious cues or extra nuisance coordinates: the optimizers separate at orders 9--11, then approach chance despite substantial remaining Bayes predictability. A mixed task establishes a recovery boundary, and CIFAR-10 supplies an external architecture comparison. Together, these findings connect optimizer comparison to the acquisition and use of specified relations, identifying permitted adaptation as a concrete training variable that changes what a fixed architecture learns.

cs.LG↗

arXiv2026

When the Preconditioning Exponent Turns Negative: Learning-Rate Coupling and Cross-Environment Generalization

Adaptive optimizers are commonly parameterized by a fixed power of the second-moment estimate. Existing partially adaptive methods study exponents between momentum-like updates and the standard Adam square root, while the interaction between this exponent and the global learning rate is less understood. We perform a controlled cross-environment study using a paired four-environment classification problem with stable sparse features, environment-dependent spurious sparse features, dense features, and high-dimensional noise. Across \NumRuns{} source-training runs covering 21 preconditioning exponents $p\in[-0.5,0.5]$ and five learning rates $η\in[10^{-4},10^{-2}]$, we find that the exponent maximizing cross-environment accuracy decreases almost linearly with $\log_{10}η$. The fitted slopes range from $-0.270$ to $-0.300$, with $R^2$ between $0.972$ and $0.996$. At $η=10^{-2}$, source-validation selection still prefers positive exponents in all four environments, whereas cross-environment and worst-environment criteria prefer negative exponents. Checkpoint decomposition shows that lower $p$ reduces the learned spurious-to-stable and noise-to-stable weight ratios; under reversed correlation, it also reduces the magnitude of the harmful spurious margin. Negative $p$ is therefore not a universally optimal setting. It is a high-step-size allocation regime produced by the joint action of learning rate and preconditioning. The study also exposes a model-selection conflict: source-domain validation systematically selects a different preconditioning regime from the one that maximizes robustness to environmental change. The results are a single-seed, finite-budget mechanism study rather than a broad benchmark claim.

cs.LG↗

arXiv2025

Natural Spectral Fusion: p-Exponent Cyclic Scheduling and Early Decision-Boundary Alignment in First-Order Optimization

Spectral behaviors have been widely discussed in machine learning, yet the optimizer's own spectral bias remains unclear. We argue that first-order optimizers exhibit an intrinsic frequency preference that significantly reshapes the optimization path. To address this, we propose Natural Spectral Fusion (NSF): reframing training as controllable spectral coverage and information fusion rather than merely scaling step sizes. NSF has two core principles: treating the optimizer as a spectral controller that dynamically balances low- and high-frequency information; and periodically reweighting frequency bands at negligible cost, without modifying the model, data, or training pipeline. We realize NSF via a p-exponent extension of the second-moment term, enabling both positive and negative exponents, and implement it through cyclic scheduling. Theory and experiments show that adaptive methods emphasize low frequencies, SGD is near-neutral, and negative exponents amplify high-frequency information. Cyclic scheduling broadens spectral coverage, improves cross-band fusion, and induces early decision-boundary alignment, where accuracy improves even while loss remains high. Across multiple benchmarks, with identical learning-rate strategies and fixed hyperparameters, p-exponent cyclic scheduling consistently reduces test error and demonstrates distinct convergence behavior; on some tasks, it matches baseline accuracy with only one-quarter of the training cost. Overall, NSF reveals the optimizer's role as an active spectral controller and provides a unified, controllable, and efficient framework for first-order optimization.

cs.LG↗

arXiv2024

The Unified Balance Theory of Second-Moment Exponential Scaling Optimizers in Visual Tasks

We have identified a potential method for unifying first-order optimizers through the use of variable Second-Moment Exponential Scaling(SMES). We begin with back propagation, addressing classic phenomena such as gradient vanishing and explosion, as well as issues related to dataset sparsity, and introduce the theory of balance in optimization. Through this theory, we suggest that SGD and adaptive optimizers can be unified under a broader inference, employing variable moving exponential scaling to achieve a balanced approach within a generalized formula for first-order optimizers. We conducted tests on some classic datasets and networks to confirm the impact of different balance coefficients on the overall training process.

cs.LG↗

arXiv2023

Asymmetric Momentum: A Rethinking of Gradient Descent

Through theoretical and experimental validation, unlike all existing adaptive methods like Adam which penalize frequently-changing parameters and are only applicable to sparse gradients, we propose the simplest SGD enhanced method, Loss-Controlled Asymmetric Momentum(LCAM). By averaging the loss, we divide training process into different loss phases and using different momentum. It not only can accelerates slow-changing parameters for sparse gradients, similar to adaptive optimizers, but also can choose to accelerates frequently-changing parameters for non-sparse gradients, thus being adaptable to all types of datasets. We reinterpret the machine learning training process through the concepts of weight coupling and weight traction, and experimentally validate that weights have directional specificity, which are correlated with the specificity of the dataset. Thus interestingly, we observe that in non-sparse gradients, frequently-changing parameters should actually be accelerated, which is completely opposite to traditional adaptive perspectives. Compared to traditional SGD with momentum, this algorithm separates the weights without additional computational costs. It is noteworthy that this method relies on the network's ability to extract complex features. We primarily use Wide Residual Networks for our research, employing the classic datasets Cifar10 and Cifar100 to test the ability for feature separation and conclude phenomena that are much more important than just accuracy rates. Finally, compared to classic SGD tuning methods, while using WRN on these two datasets and with nearly half the training epochs, we achieve equal or better test accuracy.

cs.LG↗