Image Classification with Cifar 10

TL;DR

What worked:

• Deeper networks: Yes, but too Deep a network leads to overfitting and taking very long to train
• Wider networks: Yes, try to make your networks Deeper in the middle as more and more information is captured in those layers
• Learning Rate: Try starting to train with a small learning rate(1e-5). If training is too slow, you can experiment with increasing learning rates
• Monitor Validation Metric: Helps to identify if the network is under-fitted or overfitted

Cifar 10 Description:

Given an image, classify it into one of the 10 classes

Input

Image with 3 channels(R,G,B)

Output

Cifar 10 classes: Airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck

Created Network

Code