Exploring augmentation with horses vs. humans

In this video on the topic heading the instructor told us

“If the validation set doesn’t have the same randomness,

then its results can fluctuate like this,

So bear in mind that you don’t just

need a broad set of images for training,

you also need them for testing or

the image augmentation won’t help you very much.”

And I don’t know why would that make a difference in the accuracy of the validation set.

I would understand if the real-world examples or use results would vary from what we get from the validation set due to its lack of variance.

I hope someone can clear this up a little bit more for me.

Thank you in advance.

Inspection of training pipeline is required when distribution of validation dataset is different from training dataset.
Image augmentation can help broaden training dataset distribution.

There are 2 cases when image augmentation is not the only solution and you need more images:

  1. Insufficient image augmentation of training data doesn’t fully cover validation data distribution. Increasing training examples / augmentation is required to take care of this.
  2. Validation set might be poorly designed. For instance, if all humans are standing the validation and augmented training images make the standing nature of human images hard to capture, validation accuracy might vary across epochs till the time the model has learnt to understand images in their augmented form to classify a human in standing form. The situation can be made better if there were more images in the validation set that look like the augmented training images.