Dongxu Zhang 2015-09-14

Last Week:

1. The accuracy of tfidf with 2 layer NN and softmax for classification is higher than 4 layer NN and softmax(79%, 75%). The performance on test set gets best result when training error is around 0.98.

2. Set small lambda of joint training only made the result not that bad. (tried lambda = 0.1)

3. Tried a simple idea on Graph prediction.

This Week:

1. Tried features for softmax. And tried smaller lambda.

2. tune the performance of transE. and check out if the idea works.

3. Try sequence supervised learning. Compare with Viterbi.