MLP Results

Latest results:

Experiment 2(a) (March 21, 2017)

In experiment 2(a), we slightly tweaked the model structure, but more importantly, we improved visualization of results.

Methodology for fully-connected MLP with L2 loss model:

• Neural net architecture: Fully-connected MLP
• Input size: 3*64*64 - 3*32*32 = 9216
• Output size: 3*32*32 = 3072
• Number of hidden layers: 2
• Hidden layer 1 has 512 units and uses ReLU activations
• Hidden layer 2 has 512 units and uses ReLU activations
• Loss function: Mean squared error
• Optimizer: adam
• Number of epochs: 20

Input (outer frame)	Model prediction (inner frame)	Correct output (inner frame)	Input + prediction	Input + correct output)

Experiment 1

The first column represents the target inner 32x32 image, wheras the second column represents the corresponding generated prediction of the inner 32x32 image using the outer frame as input.

Target	Prediction

Experiment 2(b) (March 23, 2017)

Here we tweaked the loss function to use Mean Absolute Error instead of Mean Squared Error.
Click here to visualize the results.

Experiment 2(c) (March 23, 2017)

Here we tweaked the model from Experiment 2(a) to use a sigmoid activations on the last layer, rather than a ReLU.
Click here to visualize the results.

Experiment 3 (March 26, 2017)

Methodology for GAN model:

Reference: Inspired from jacobgil's keras-dcgan implementation