If Training data is not similar to testing data ?

20 views

Skip to first unread message

unread,

Aug 22, 2016, 3:41:56 AM8/22/16

to torch7

I try to train a neural network in an emotional recognition dataset for five class. I meet some question when I trained.

The dataset set is a skewed dataset , ex : class1 has 100 datas but class2 has 3000 datas.

-- I try to use Smote to balance data

The training data of this dataset is not similar to testing data , that is a big problem

I don't know how to solve problem 2 please help me. Thanks a lot

unread,

Aug 24, 2016, 4:59:50 AM8/24/16

to torch7

Which kind of learning do you use ?

Supervised Learning ?

Unsupervised, by renforcement ??

Generally when you want to train a NN on a classification problem (your five class) you must have similarities between training and testing data..