Source author record

Qingjun Zhang

Qingjun Zhang appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

Catalog footprint

What is connected

2works
6topics
4close collaborators

Actions

Connect this record

Log in to claim

Research graph

See the researcher in context

Open full explorer

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

Published work

2 published item(s)

preprint2026arXiv

E = T*H/(O+B): A Dimensionless Control Parameter for Mixture-of-Experts Ecology

We introduce E = T*H/(O+B), a dimensionless control parameter that predicts whether Mixture-of-Experts (MoE) models will develop a healthy expert ecology or collapse into dead experts. E combines four hyperparameters -- routing temperature T, routing entropy weight H, oracle weight O, and balance weight B -- into a single quantity. Through 12 controlled experiments (8 vision, 4 language) totaling over 11,000 training epochs, we establish that E >= 0.5 alone is sufficient to guarantee zero dead experts, removing the necessity for handcrafted load-balancing auxiliary losses. We validate this cross-modally on CIFAR-10, CIFAR-100, TinyImageNet-200, WikiText-2, and WikiText-103. Six additional findings emerge: (1) dead experts can resuscitate -- triggered by balance loss driving router re-exploration; (2) ortho toxicity is dataset-dependent, not universal; (3) task complexity shifts the critical E threshold; (4) model overfitting is decoupled from expert ecological health; (5) three-tier MoE spontaneously collapses into a two-tier functional structure; (6) ecological structure is temperature-invariant across a 50x range. We propose that E serves as a unified diagnostic for MoE training, analogous to the Reynolds number in fluid dynamics.

preprint2012arXiv

Evaluation of Scale-Invariance In Physiological Signals By Means Of Balanced Estimation Of Diffusion Entropy

By means of the concept of balanced estimation of diffusion entropy we evaluate reliable scale-invariance embedded in different sleep stages and stride records. Segments corresponding to Wake, light sleep, REM, and deep sleep stages are extracted from long-term EEG signals. For each stage the scaling value distributes in a considerable wide range, which tell us that the scaling behavior is subject- and sleep cycle- dependent. The average of the scaling exponent values for wake segments is almost the same with that for REM segments ($\sim 0.8$). Wake and REM stages have significant high value of average scaling exponent, compared with that for light sleep stages ($\sim 0.7$). For the stride series, the original diffusion entropy (DE) and balanced estimation of diffusion entropy (BEDE) give almost the same results for de-trended series. Evolutions of local scaling invariance show that the physiological states change abruptly, though in the experiments great efforts have been done to keep conditions unchanged. Global behaviors of a single physiological signal may lose rich information on physiological states. Methodologically, BEDE can evaluate with considerable precision scale-invariance in very short time series ($\sim 10^2$), while the original DE method sometimes may underestimate scale-invariance exponents or even fail in detecting scale-invariant behavior. The BEDE method is sensitive to trends in time series. Existence of trend may leads to a unreasonable high value of scaling exponent, and consequent mistake conclusions.