{"resolved":true,"match_type":"uuid","source_type":"original_book","book_id":"cf76bc49-7e9b-4f42-bee8-3050a9fb9feb","title":"智慧的疆界 从图灵机到人工智能","author":null,"edition":null,"publisher":null,"year":null,"isbn":"9787111778004","requested_by":"printed_page","book":{"source_type":"original_book","book_id":"cf76bc49-7e9b-4f42-bee8-3050a9fb9feb","id":"cf76bc49-7e9b-4f42-bee8-3050a9fb9feb","uuid":"cf76bc49-7e9b-4f42-bee8-3050a9fb9feb","title":"智慧的疆界 从图灵机到人工智能","normalized_title":"智慧的疆界从图灵机到人工智能","author":null,"edition":null,"publisher":null,"year":null,"isbn":"9787111778004","isbns":["9787111778004","9787111695714"],"parsed":true,"ai_readable":true,"document_type":"book","page_count":456,"navigation_url":"https://workshop.closeclaw.top/ai/book/cf76bc49-7e9b-4f42-bee8-3050a9fb9feb","title_url":"https://workshop.closeclaw.top/ai/title/%E6%99%BA%E6%85%A7%E7%9A%84%E7%96%86%E7%95%8C%20%E4%BB%8E%E5%9B%BE%E7%81%B5%E6%9C%BA%E5%88%B0%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD","fixed_page_url_template":"https://workshop.closeclaw.top/ai/book/cf76bc49-7e9b-4f42-bee8-3050a9fb9feb/page/{printed_page}","path_api_url_template":"https://workshop.closeclaw.top/api/ai/book/cf76bc49-7e9b-4f42-bee8-3050a9fb9feb/page/{printed_page}","read_url":"https://workshop.closeclaw.top/api/ai/read/cf76bc49-7e9b-4f42-bee8-3050a9fb9feb","page_url":"https://workshop.closeclaw.top/api/ai/books/cf76bc49-7e9b-4f42-bee8-3050a9fb9feb/page","search_url":"https://workshop.closeclaw.top/api/ai/books/cf76bc49-7e9b-4f42-bee8-3050a9fb9feb/search"},"page":{"printed_page":316,"pdf_page":331,"physical_index":331,"physical_index_base":0,"printed_page_status":"recognized","page_sha256":"b2136fe3257fdfeb2f9604dd40674c5dacbfbf55a8ffb229ac935addba2e19d6"},"chapter":"内容简介","content":"为了便于读者理解，笔者借助一个虚构示例来解释古德费洛的设计思路。以印假钞为例，干这种买卖最大的竞争对手并非同行，而是警察。如果假钞能够精致到连警察也无法分辨的程度，那就算完全掌握了“核心竞争力”。而作为警察，职责自然是持续提升自己的装备和业务水平，以便能够准确地分辨出真钞与假钞。在罪犯与警察造假鉴真的对抗游戏里面，就隐藏了生成式模型突破的奥秘。\n\n首先，请读者扮演警察的角色，允许使用任何手段来识别一张图像（譬如图 8-2）是否真钞。目前效果最好的方法自然是使用基于深度学习训练的判别式模型，判别效果能好到什么程度呢？我们可以参考第 7 章中 ImageNet 大规模视觉识别挑战赛的结论，把 ILSVRC 数据集交由人类来判别的平均错误率是 5.1%，而 2015 年的冠军—残差网络（ResNet）的判别错误率已经达到了3.57%，已超越了人类平均水平。换言之，此时的技术水平已经可以让计算机扮演一个判别能力和人类平均水平大致相当的警察了。\n\n接下来，请读者扮演罪犯的角色。一开始你印假钞的手段生疏拙劣，生成的图像几乎与完全随机的噪声无异，警察一眼就能看出真假。不过，只要警察将判别的结果告知你，你就有了强化学习的依据，知道了改进造假手段的方向。得到来自警察的反馈，你痛定思痛并经过长期练习，手法逐渐成熟，生成的结果逐渐从随机变得有序，也逐渐有了一些能以假乱真逃过警察法眼的高质量图像。\n\n316\n\n第三部分 第三波高潮","previous_page":315,"next_page":317}