知识卡片
跨层降维需要说明共同坐标
原书把所示层的表示联合送入t-SNE,使投影共享一次拟合的坐标系。它提供跨层观察线索,但t-SNE主要强调邻近关系,不能据二维图断言全局距离成比例保持。若各层单独降维,任意旋转和布局差异又会进一步妨碍直接比较。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
colors) are separated. This deep net is a vision transformer (ViT [109]), which we will learn about in section 26.7. It was trained using contrastive language-image pre-training (CLIP [394], see section 51.3). Each ViT block contains multiple layers of neural processing (see figure 26.14; we visualize the embeddings right after the first token norm in a block). We apply t-SNE jointly across all shown layers.
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第13章《作为分布变换器的神经网络》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c013.xhtml)
- 结论依据:摘录中关于跨层降维需要说明共同坐标的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:colors) are separated. This deep net is a vision transformer (ViT [109]), which we will learn about in section 26.7. It was trained using contrastive language-image pre-training (CLIP [394], see section 51.3). Each ViT block contains multiple layers of neural processing (see figure 26.14; we visualize the embeddings right after the first token norm in a block). We apply t-SNE jointly across all shown layers.