Supervised Learning - Classification Week 9 Challenge

1-What is a Neural Network?

It will work just like our brain works and get trained from the signals passed by n-number of neurons(node).

It has multiple neurons as a hidden layers which receives the multiple input feature and predicts/classifies the output.

Basic working of NN:

Working of NN can be explained in 2 steps: i)-Forward Propagation, ii)-Backward Propagation.

i)-Forward Propagation:Input features with particular weights are passed to the neuron, within a neuron i)-Y value would be calculated,ii)-Activation function were applied to y (Like sigmoid activation function) decides whethwer the output can be classified or not and the output would be predicted.

ii)-Backward Propagation:

Once the output has been predicted by the neuron, the cost for prediction were calculated. If the loss is not reduced, will use optimer function(like gradient descent, stochastic gradient) to change the weights by applying optimized functions such as Gradient decent (the derivative with loss and learning rate). After changing the weights will do the forward propagation again and check for loss. This process would be carried out untill there is minimal loss value.

2)-What is deep learning?

Deep learning mimcs the working of our brain. This is a subset of Machine Learning. Peforms the prediction based on one or multiple layers of Neural Networks to make approximate predictions.

3)-What is the role of Activation Function?

Activation function plays a major role in NN. It intimate the neuron for classifing the output. Basically it decides whether the neuron need to get activated or de-activated for output.

Without activation function, NN would have the linear relation between weights and bias. Due to this we cannot perform complex problems.

Activation function provides a non-linear relation to the input and make it capable to perform complex classifications. Activation function uses differential functions which provides a easy way to perform back propagation.

4)-What are the different types of Activation functions?

i>Rectified Linear Unit: This activation function ranges from 0-1(all -ve values are considered 0) which says it can map or fit the data properly.

ii)-Leaky ReLU Function: In this function, the range has been extended for removing the unfitting issue in ReLU. Not all the -ve values are considered 0,but the values near to 0.

iii)-Sigmoid: This function is used in Logistic Regression which is a Probalistic approach results the value range from 0-1. Most of the NN models would be considering this sigmoid function since it has minimal range and it is based on prediction which would be more accurate.

iv)-Hyperbolic Tangent: This is similar to sigmoid function, but the negative values would be considered and range is from -1 to 1.

5)-Backward Propagation:

Once the output has been predicted by the neuron, the cost for prediction were calculated. If the loss is not reduced, will use optimer function(like gradient descent, stochastic gradient) to change the weights by applying the derivative with loss and learning rate. After changing the weights will do the forward propagation again and check for loss. This process would be carried out untill there is minimal loss value.