通用Transformer（Universal Transformer，UT）是2018年谷歌研究提出的架构，旨在通过引入循环机制增强标准Transformer处理长序列和动态计算的能力。其核心思想是用共享的循环过渡函数替换固定层数的前馈网络，使模型能根据输入自适应地进行推理步骤。以下是其技术要素的可视化深度解析（文本描述形式）： **1. 核心问题：标准Transformer的固定深度限制** 在标准Transformer中，每个token的表示通过固定数量（如L=6）的相同结构层（自注意力+前馈网络）进行变换。这种固定深度导致： - 对于简单token可能过度计算（浪费） -

Prompt

Create a 10-second video of a miniature LEGO city traffic scene with tiny cars and pedestrians moving through a busy plaza. Use plastic shine, modular block geometry, a cinematic tracking shot, bright daylight, and a playful energetic mood.

Create a 10-second shadow puppet performance of a bird flying across the moon and transforming into a dragon. Use high-contrast silhouettes, a stage curtain background, theatrical lighting, and elegant slow motion.

Create a 10-second watercolor animation of a girl running through a field of wildflowers in the rain. Use soft bleeding pigments, painted motion, a gentle pan, and a dreamy nostalgic atmosphere.

Create a 10-second comic book action scene of a hero leaping across rooftops in a rainy city. Use bold ink outlines, halftone texture, a dramatic whip-pan, and a high-energy mood.