神经网络数量多时如何使用 类
How to use a neural network when there are many classes
我们正在使用 facenet 并为面部 https://github.com/davidsandberg/facenet. We have 100k classes (celebrities) from MSCeleb http://www.msceleb.org/ 和 800 万个样本生成了嵌入（128 个特征）。
如何构建一个可以将 128 个特征映射到 100k 类 的神经网络？
使用全连接层会导致 (128 + 1)*100k = 1290 万个参数，这似乎太大而无法训练。
来自 FaceNet 摘要：
In this paper we present a system, called FaceNet, that directly
learns a mapping from face images to a compact Euclidean space where
distances directly correspond to a measure of face similarity. Once
this space has been produced, tasks such as face recognition,
verification and clustering can be easily implemented using standard
techniques with FaceNet embeddings as feature vectors.
与其训练分类器，不如考虑在特征中进行最近邻搜索 space。您可以 select 为 10 万名名人中的每一个锚定图像，然后根据他们的特征向量构建 k-d tree。然后对于每个输入，你可以在 k-d 树中找到它最近的邻居。
我们正在使用 facenet 并为面部 https://github.com/davidsandberg/facenet. We have 100k classes (celebrities) from MSCeleb http://www.msceleb.org/ 和 800 万个样本生成了嵌入（128 个特征）。
如何构建一个可以将 128 个特征映射到 100k 类 的神经网络？
使用全连接层会导致 (128 + 1)*100k = 1290 万个参数，这似乎太大而无法训练。
来自 FaceNet 摘要：
In this paper we present a system, called FaceNet, that directly learns a mapping from face images to a compact Euclidean space where distances directly correspond to a measure of face similarity. Once this space has been produced, tasks such as face recognition, verification and clustering can be easily implemented using standard techniques with FaceNet embeddings as feature vectors.
与其训练分类器，不如考虑在特征中进行最近邻搜索 space。您可以 select 为 10 万名名人中的每一个锚定图像，然后根据他们的特征向量构建 k-d tree。然后对于每个输入，你可以在 k-d 树中找到它最近的邻居。