“Wasn't it amazing that they could distill the concept of cat from images with no help from external labels”
Building high-level features using large scale unsupervised learning
Google 用 1.6 万核无监督训练,网络自发学出人脸和猫脸检测器
论文探讨能否仅用无标注数据学出类别特定的高层特征检测器:训练一个 9 层的局部连接稀疏自编码器(含池化和局部对比度归一化),拥有 10 亿连接,数据为 1000 万张网络图像,在 1000 台机器(1.6 万核)上用模型并行与异步 SGD 训练三天。结果表明无需标注也能得到人脸检测器,对平移、缩放和平面外旋转都有鲁棒性,还对猫脸和人体敏感;以这些特征为起点,在 20000 类的 ImageNet 识别上达到 15.8% 准确率,相对此前最优提升 70%。
HN 上的人把它当作“Google 的猫脸实验”的出处:网络没有任何标签就学出了猫这个概念,甚至有人因此决定转行读博做 AI 研究。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“I did only print it (yet), but from the abstract it appears that they did mainly tried face recognition, not trying to find cats.”
“Paper here: Building high-level features using large scale unsupervised learning”
“Given the 15.8% success rate of the cat detector on large data sets: Building high-level features using large scale unsupervised learning,”
“Unsupervised image feature extraction[1]”
“Fast forward to 2012, when I saw this paper discussed in the news: Building high-level features using large scale unsupervised learning - that was when I decided to quit my job and enroll in a phd program to do AI research.”
- Google research (with Jeff Dean): Face recognition with unsupervised learning ↗tonfa · 2012-05-30 · 3 分