Yuwen Hao

Jinchang Zhu, Jindong Li, Yuwen Hao et al.

A causal-decoder block is hierarchical: lower layers build the residual basis that upper layers attend over. We identify a failure mode in GPT pretraining: upper layers commit to sharp attention patterns before lower-layer features stabilize. We call this premature upper-layer attention specialization. Temporarily slowing only upper-layer Q/K projections during early training improves final perplexity and downstream accuracy without altering other parameters; it prevents upper attention from collapsing onto an immature residual basis. In LLaMA-style blocks, the same intervention is nearly unnecessary. Through ablations, we isolate multiplicative gated FFNs (not RMSNorm or bias removal) as the component that suppresses the upstream residual writes driving the failure. A pathwise analysis unifies both findings: the learning-rate intervention reduces a step-size factor, while gated FFNs reduce a residual-energy factor on the same growth pathway. Our results identify upper-layer Q/K timing as a concrete interaction point between decoder architecture and optimization.

Chuang Wang, Zhengping Li, Yuwen Hao et al.

In order to solve the problems of long training time, large consumption of computing resources and huge parameter amount of GAN network in image generation, this paper proposes an improved GAN network model, which is named Faster Projected GAN, based on Projected GAN. The proposed network is mainly focuses on the improvement of generator of Projected GAN. By introducing depth separable convolution (DSC), the number of parameters of the Projected GAN is reduced, the training speed is accelerated, and memory is saved. Experimental results show that on ffhq-1k, art-painting, Landscape and other few-shot image datasets, a 20% speed increase and a 15% memory saving are achieved. At the same time, FID loss is less or no loss, and the amount of model parameters is better controlled. At the same time, significant training speed improvement has been achieved in the small sample image generation task of special scenes such as earthquake scenes with few public datasets.