{"resolved":true,"match_type":"uuid","source_type":"original_book","book_id":"cd4f7678-cd9b-4971-bde0-e9ad46a6cc71","title":"人工智能与大数据技术导论","author":null,"edition":null,"publisher":null,"year":null,"isbn":"9787302517986","requested_by":"printed_page","book":{"source_type":"original_book","book_id":"cd4f7678-cd9b-4971-bde0-e9ad46a6cc71","id":"cd4f7678-cd9b-4971-bde0-e9ad46a6cc71","uuid":"cd4f7678-cd9b-4971-bde0-e9ad46a6cc71","title":"人工智能与大数据技术导论","normalized_title":"人工智能与大数据技术导论","author":null,"edition":null,"publisher":null,"year":null,"isbn":"9787302517986","isbns":["9787302517986"],"parsed":true,"ai_readable":true,"document_type":"book","page_count":490,"navigation_url":"https://workshop.closeclaw.top/ai/book/cd4f7678-cd9b-4971-bde0-e9ad46a6cc71","title_url":"https://workshop.closeclaw.top/ai/title/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD%E4%B8%8E%E5%A4%A7%E6%95%B0%E6%8D%AE%E6%8A%80%E6%9C%AF%E5%AF%BC%E8%AE%BA","fixed_page_url_template":"https://workshop.closeclaw.top/ai/book/cd4f7678-cd9b-4971-bde0-e9ad46a6cc71/page/{printed_page}","path_api_url_template":"https://workshop.closeclaw.top/api/ai/book/cd4f7678-cd9b-4971-bde0-e9ad46a6cc71/page/{printed_page}","read_url":"https://workshop.closeclaw.top/api/ai/read/cd4f7678-cd9b-4971-bde0-e9ad46a6cc71","page_url":"https://workshop.closeclaw.top/api/ai/books/cd4f7678-cd9b-4971-bde0-e9ad46a6cc71/page","search_url":"https://workshop.closeclaw.top/api/ai/books/cd4f7678-cd9b-4971-bde0-e9ad46a6cc71/search"},"page":{"printed_page":343,"pdf_page":359,"physical_index":359,"physical_index_base":0,"printed_page_status":"recognized","page_sha256":"43e311dde4d7e633b1f90eba5cf9f10799a60799ce1cdceb18aa18953ca58b1c"},"chapter":"第10章","content":"不能使用Softmax。\n\n必须依赖多个逻辑回归。\n\n例如，假设样本是只包含一项内容（一块水果）的图片。Softmax可以确定该内容是梨、橙子、苹果等的概率。如果样本是包含各种各样内容（几碗不同种类的水果）的图片，那么必须改用多个逻辑回归。\n\n总之，在训练多类别分类时，我们有几个选项可以选择。一个是使用完整Softmax，此时训练成本相对昂贵。想象一下，如果有一百万个类别，那么基本上需要为每个示例分别训练一百万个输出节点。我们可以通过进行“候选采样”来提高一点效率。\n\n10.3.3代码实例\n\n本节对手写数字进行分类。我们训练线性模型和神经网络，对传统MNIST数据集中的手写数字进行分类，并比较线性分类模型和神经网络分类模型的效果，最后可视化神经网络隐藏层的权重。我们的目标是将每个输入图片与正确的数字相对应。我们会创建一个包含几个隐藏层的神经网络，并在顶部放置一个归一化指数层，以选出最合适的类别。下面来看一下代码。\n\n首先，下载数据集，导入TensorFlow和其他实用工具，并将数据加载到Pandas DataFrame。此数据是原始MNIST训练数据的样本，我们随机选择了10000行。\n\n对这10000行数据进行训练：\n\n第10章 神经网络\n\n343","previous_page":342,"next_page":344}