Artificial Neural Network:
Deep Learning:
Deep learning is a subfield of machine learning concerned with algorithm inspired by the structure and function of the brain called “Artificial Neural Network”.
There are major contributions from deep learning, and below are few applications:
- Fraud detection
- Customer relationship management systems
- Computer vision
- Vocal AI
- Natural language processing
- Data refining
- Autonomous vehicles
- Supercomputers
- Investment modeling
- E-commerce
- Emotional intelligence
- Entertainment
- Advertising
- Manufacturing
- Healthcare
Anatomy of Neuron:
In a nutshell, below is the function of a neuron.
Dendrites: Used as an input receivers
Cell Body: Used as Transformation, Memory
Axon: is a stem for processing output.
Synapse: Is the output from the neuron
Perceptron:
Perceptron's work in a similar way like Neuron, it takes input and perform transformations and produces the results.
Inside the perceptron, we typically calculate the step function. Step functions, it will take the cumulative sum of weights and inputs, along with the bias.
If step function output is greater than 0, then its a fire, otherwise it’s a dormant. Fire is to move further/trigger, dormant is like produced the result straight way.
There is Multi Layer Perceptron, where it mimic same as perceptron, except the function introduce and the function is a non-linear function.
A perceptron has the following components:
• Input layer
- Weights
- Bias
- Net input function
- Activation function
Mathematical way of understanding perceptron:
From the above representation, there are inputs (x1) and associated weights, will be calculated the cumulative summation with bias (w1x1+w2x1+w3x1+w4x1+b). And if the problem is of the classification we can assume the output Y is +1 or -1, So from the output of the summation and the step function/Activate function is >0, then it classify to +1, if the output is <0, then to classify to -1.
Question comes to mind is what is the criterion to choose weights and bias, and what are the valid separator's. Choosing the weights and bias are arbitrary, where taking the weights and bias randomly and adding the error term and misclassified datapoint.
Example:
lets take w1, as an arbitrary initialization, and the expression is
w1+1 < == w1+1 + y1+1*x1+1
W1+2 < == w1+2 + y1+2*x1+2
where y1+2*x1+2 is an error term and this becomes this misclassified datapoint. And that means it’s not an valid separator.
Artificial Neural Network:
Artificial Neural Network, is a form with a multiple perceptron’s. Where input will taken by the another perceptron. And in the artificial neural network, activation function will be added along with the weights as an input. In a nutshell, Activation function is an output of a every neuron, and will be added as a cumulative sum of activation and weighted sum as an input.
There are six main things that need to be specified for specifying a neural network completely:
- Network Topology
- Input Layer
- Output Layer
- Weights
- Activation functions
- Biases
