How would I be able to improve my CNN model (Keras)?

Question

Recently I read a research paper on age detection using facial images. So right now because of that I was trying to see how far I could get by applying a CNN to a dataset of facial images (with their respective ages) in order to predict their ages which would be in bins (ex. 0-10, 11-20, 21-30...).

For training and testing

training.shape (50000, 28, 28)
testing.shape (2938, 28, 28)

I tried to keep the images small as they would be able to run faster as well as using grayscale. And for the actual layers themselves I tried to keep it simple, for now,

model = Sequential()
model.add(Conv2D(64, kernel_size=3, activation='relu', input_shape=(28,28,1)))
model.add(Conv2D(32, kernel_size=3, activation='relu'))
model.add(Flatten())
model.add(Dense(10, activation='softmax'))


Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 26, 26, 64)        640       
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 24, 24, 32)        18464     
_________________________________________________________________
flatten_1 (Flatten)          (None, 18432)             0         
_________________________________________________________________
dense_1 (Dense)              (None, 10)                184330    
=================================================================
Total params: 203,434
Trainable params: 203,434
Non-trainable params: 0
_________________________________________________________________

Compiled with the following

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

So far the best accuracy after running for 100 epochs has been 37.16. Which isn't great but recently I've gotten access to one of my schools gpu's so I wanted to fix anything that I'm doing wrong and improve my model. Is there anything you could recommend when it comes to improving the model, theres probably a lot this is more my first time trying to do this.

Answer 1

The first thing you should notice is that you've almost ruined your input signal. Take a look at a $28\times28$ image of a face? what can you see? is there any difference between a teenager and a middle-aged person? The point is that the network should be trained using data that does not have high Bayes error which means you as an expert can distinguish between inputs and label them correctly. Increase the size of your inputs. By doing so, if you use the current regime, you may have lots of trainable parameters between dense layers and convolutional layers. Consequently, try to employ more convolutional layers and some pooling layers among them. Also, try to add more dense layers with more neurons in each.

Answer 2

Two things come to my mind that you can try quickly.

1. Do NOT shrink your images, 28x28 is too small (it is fine for MNIST-like datasets, but not for faces). Also by making them grayscale you are missing a lot of valuable information, try to use color images.
2. Use a pretrained CNN, keras offers a number of them, I normally play quite a bit with VGG16 as it is a simple network to reuse. My recommendation is to freeze all the layers but the last one and see which performance you get (as a baseline). Then considering unfreezing other layers for increased performance.

Please do give it a shot to those options, as per the GPU, it is pretty much a MUST for CNNs, notice that kaggle.com is providing now jupyter notebooks with FREE GPU (it is on beta atm, but seems to do the job quite well).