Transformer 注意力机制可视化

生成一张 4:5 竖版「Transformer 注意力机制可视化」案例图，所属分类为「科研与学术图」。Transformer 双头注意力热图，8×8 token 方格，紫色深度编码，语法依存箭头标注，上方 Q/K/V 方块简图。画面需要完整呈现上述主体、构图、配色、材质、光线和整体风格；可见文字以自然简体中文为主，必要的技术缩写、代码、变量名、学术符号、单位、拉丁物种名或装饰性外文可保留。

A minimalist scientific visualization of multi-head self-attention in a transformer model, displayed as an annotated attention heatmap. The heatmap is a square grid (8×8 tokens) where rows represent query tokens and columns represent key tokens; cell color intensity encodes attention weight from white (zero) to deep violet (maximum). Token labels along both axes read: "The", "cat", "sat", "on", "the", "mat", ".", "[EOS]" in 8 pt monospace. Two attention heads are shown side by side: Head 1 heatmap on the left, Head 2 heatmap on the right, each with its own color scale bar below in matching violet-to-white gradient. A thin annotation arrow on Head 1 points to the strong "cat" → "sat" attention cell, labeled "syntactic dependency" in 7 pt sans-serif. Above the heatmaps, a simplified transformer block schematic (Q/K/V rectangles, softmax arrow, output) is shown in flat pale-blue. White background, monochrome violet palette for attention weights.