Deep Regression Networks
data:
x: 1x44 vector
y: 1x2 vector
using mlp fitting,
the loss function mean(abs(pred-target)/abs(target))
the best result is 5% relative error
note:
- learning rate is import to this task, and we set learningRate=0.04 which is the best
- drop rate can't be too large, which may be harmful to the model, since the input dimension is only 44,in this task we using dropout rate=0.001
3.activation functions of relu may not help to improve reduce the loss - more data is good for this task