A simple implementation of a deep neural network to answer the most important question of all ...
Is this a cat?
The cat's name is Lisa. She is adorbable and belongs to my girlfriend's mother. We implement and compare both logistic regression and least squares regression for the cost function. The network has the following architecture:
- 1 input layer (with 12288 nodes)
- 3 hidden layers (with 20, 7, and 5 nodes; respectively)
- 1 output layer (with 1 node)
The input layer has 12288 nodes because the images are 64 pixels x 64 pixels and contain 3 RGB color layers (i.e. 64 x 64 x 3 = 12288).
The network correctly predicts the above photo as a 'cat' picture. Example output:
y = 1.0, your model predicts a "cat" picture"
The network uses batch learning. Here's an illustration of the performance over repeated iterations:
Reduction in cost (a) when using logistic regression; (b) when using least squares regression
Since this is a classification problem (just 1's and 0's), logistic regression works better (note the faster reduction in cost above). The network makes some mistakes. Here is a print out of the incorrectly classified images when using logistic regression:
Note: this code was developed in partial fufillment of the Deep Learning Specialization program under Andrew Ng et al.