课程定位
CS 236G 是生成模型专题课,聚焦 GAN。课程从最基础的对抗训练讲到 StyleGAN 等 SOTA 形式,讨论模式崩溃、训练稳定性、评估指标和偏见,也覆盖图像编辑、医学低数据、气候可视化等应用。
主要内容
- GAN 基础:生成器/判别器、minimax、损失函数
- 训练技巧:WGAN、谱归一化、TTUR、正则
- 条件生成:cGAN、StyleGAN、图像到图像
- 评估:FID、IS、精确率/召回率
- 偏见与伦理:数据偏差、深伪
- 应用:编辑、医学、科学可视化
- 项目
先修
线性代数、统计、CS 106B;研究生级 AI 课(CS 229/230/221)。
考核与工作量
作业 + 项目。具体以当季官网为准。
适合谁
图像生成、视觉编辑、生成模型工程;想深入理解 GAN 而不是只会调 diffusion 的人。
学习建议
先修 CS 236 或 CS 231N。GAN 的训练稳定性问题最好亲手复现一次。
相关课程
官方课程描述(英文原文)
Generative Adversarial Networks (GANs) have rapidly emerged as the state-of-the-art technique in realistic image generation. This course presents theoretical intuition and practical knowledge on GANs, from their simplest to their state-of-the-art forms. Their benefits and applications span realistic image editing that is omnipresent in popular app filters, enabling tumor classification under low data schemes in medicine, and visualizing realistic scenarios of climate change destruction. This course also examines key challenges of GANs today, including reliable evaluation, inherent biases, and training stability. After this course, students should be familiar with GANs and the broader generative models and machine learning contexts in which these models are situated. Prerequisites: linear algebra, statistics, CS106B, plus a graduate-level AI course such as: CS230, CS229 (or CS129), or CS221.