New Idea For Today
We can expand our previous PCA approach to an autoencoder, which converts a time series to specific output, just like what we did in last blog, which uses LSTM to encode & decode. Then calculate reconstruction error.
What is different is to start a new direction:
- like what I thought about in simple application as a word generation machine, can add another layer as encoder/decoder
- Also, we can use encoder as novelty detection tool, just as we do in last paper
- also, other novel applications, like huffman tree, word2vec(which I really want to apply to), more efficient Brown cluster etc.
this one can only be limited to low dimensional things. For high dimensional data, we might think about a generative approach: P(s1, s2, s3 …) as joint probability.