Chao Xing 2016-01-25

Last Week:

---

Solution:

• None.

Plan to do:

   1. Test DRAW in transfer large scale image style.
   2. Code basic experiments in balance word frequency.
   3. Find some unsupervised strategies in word sense vector.
   4. Test knowledge matrix factorization.

This Week:

---

What have done:

   1. Test DRAW in transfer large scale image style.
      1. LSTM for reconstruction images.
      2. LSTM add random property :
            Results : normal distribution failure to reconstruct images.
                      prior distribution mean = 0, std = 0.1 reconstruct images but lost random property.
                      prior distribution mean = 0, std = 0.25 contain both reconstruction and random property. But the random property is just change a little edges of images.
                      Consider the cost function contain both reconstruction error and KL-Divergence between prior distribution and training distribution.
                      Lost = reconstruction + 0.25 * KL. The images shows clearly, and have some random property in edges.
                      Lost = reconstruction + 0.5 * KL. Some of images also shows clearly but others are fuzzy. 
   2. Code basic experiments in balance word frequency.
   

Plan to do:

   1. Test DRAW in transfer large scale image style.

   2. Prepare technique report in combine tf-idf vector as low frequency words and distributional vector as high frequency words in text classification and sentiment classification.