Over-fitting of model with SMART Data

[e-baumer]

The model is over-fitting with MERON SMART data. Testing loss is decreasing; however, validation loss is not. This is probably related to fact that we only have ~4k images but have 4096 features. This worked for MORPH data which had ~60k images.

We should reduce the number of input features. Since the the input features come from VGG pre-trained model we need to think of a way to reduce the dimensionality. Perhaps an auto-encoder.