Dear Rajeevan Arunthavanathan ,

Look the link, maybe helpful.

http://thalis.math.upatras.gr/~vrahatis/papers/other/EpitropakisV05_ACM_SIGSAM_B_39_pp118-121_2005.pdf

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.623.9541&rep=rep1&type=pdf

Regards,

Shafagat

You could use the following techniques:

1) Visualization Activation: Yosinski et al, “Understanding Neural Networks Through Deep Visualization”, ICML DL Workshop 2014.

2) Maximally Activation patchs : Zeiler and Fergus, “Visualizing and Understanding Convolutional Networks”, ECCV 2014

3) Saliency maps and Gradient Ascent : Simonyan, Vedaldi, and Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification Models

and Saliency Maps”, ICLR Workshop 2014.

There is no free lunch theorem for this but recently a paper published which can predict the performance of another neural network prior to training (TAPAS) developed by IBM's artificial intelligence Lab. You can check the paper entitled "TAPAS: Train less accuracy predictor for architecture search". I hope this will help.