Error And Bias

http://www.ilo.org/public/english/bureau/stat/download/cpi/ch11.pdf

http://www.mathworks.com/matlabcentral/answers/110399-difference-between-true-and-apparent-error-in-neural-network

There are several quantities that can be identified.

 1. The underlying error-free I/O transformation 
 2. The contaminated sample (noise, interference and measurement error) from which the neural network is designed.
 3. The training, validation and test set targets and outputs.

The true error, which is unknown, is the difference between 1 and 2.

The true bias is the average of the true error.

The apparent error is the difference between the test set target and output.

The apparent bias is the average of the test set error.

For a good minimum mean-squared-error design, the test set bias should be zero and the test set variance is the mean-squared-error.

Hope this helps.