How to train a Binary Deep neural network for image detection.
##Summary This experiment serves as a tutorial on building a deep neural network model using Azure ML. We will be using an small image data set and build a model for predicting if a given image correspond to a human or not. ##Description ##Data From ARGB images of 64x64 pixels, we take an array of 16384 parameters. We use two dataset, one for train with 300 images and one for testing of 100. Then the datasets are arrays of size 16384x300 (train) and 16384x100 (test). The values are the pixel value normalized (value/255) and takes values form 0 to 1. ##Model We have a fully connected deep learning model that haves three hidden layers: - Input layer of 16384 nodes (the input parameters). - Hidden layer of 100 nodes with the ReLu activation function. - Hidden layer of 50 nodes with the ReLu activation function. - Hidden layer of 5 nodes with the ReLu activation function. - Binary Output layer with the sigmoid activation function. You can modify the configuration of the layer at your own to get different results. ##Algorithm Selection I used a Binary Neural Network and to construct the hidden layers and activation function, I used the #NET language. ##Results Even the accuracy is not high (the train dataset is very low), but the comparison between trained and test data is enough. Users can take this experiment and tweak the parameters of either algorithms to achieve higher performance.