这三个网络没有经过训练：没有训练数据，没有一步梯度下降。每一个权重都由一个强大的语言模型直接编写。模型 2 和模型 3 是在前一版基础上重新编辑出来的。 None of these networks was trained. There was no training data and not a single gradient step. A strong language model wrote every weight directly. Models 2 and 3 are re-edited versions of the one before.
01 · 参数量Parameters
参数量按完整的稠密矩阵计算。因为每个权重都是有意写下的，绝大多数位置都是 0——真正起作用的只有不到 1%。Parameters are counted as full dense matrices. Because every weight was placed on purpose, nearly every entry is zero: fewer than 1% of them do anything.
02 · 层数Layers
从上到下是一句话流过网络的顺序。编号的是带权重的层；虚线框没有权重。彩色的是这一版新增的层，条的长度表示该层的单元数（对数刻度）。Read top to bottom: the order a sentence flows through the network. Numbered boxes are layers with weights; dashed boxes have none. Colored boxes are new in that model. The bar under each layer shows how many units it has (log scale).
模型 2 和模型 3 的全连接层与前面所有层直接相连（稠密跳连）。In Models 2 and 3, every dense layer connects to all layers before it (dense skip connections).
03 · 区别Differences
三个模型读的都是关键词，不是语法。区别在于它们能把关键词组合成多复杂的判断。All three read keywords, not grammar. They differ in how much they can do with those keywords.
以下是三个模型对这些句子的真实输出。These are the models' real answers to these sentences.
04 · 基准测试Benchmark
模型 1 很少说"真"，所以几乎不误报，但漏掉了大部分真陈述。模型 3 找回了更多真陈述，代价是多了一些误报。Model 1 rarely says "true", so it almost never raises a false alarm, but it misses most true statements. Model 3 finds far more true statements, at the cost of more false alarms.