Easiest Explanation For Neural Networks In Artificial Intelligence. -BASICS

 Neural Networks

How do we define a Neural Network?

It can be closely related to our neuron network which works in our body. Its model is inspired the way our neurons works and creates signals to transmit useful information to cell body and then to other responsive organs.

According to Artificial Intelligence - we can understand Neural Networks as an computational system made up of Simple Highly interconnected processing elements, which process information by their dynamic state and response to external inputs.

- The basic computational Unit of a Neural Network is (Node).

- It receives input from other nodes or from external source and pass it to another node for computation of Output.

- It input has an associated Weight(W) which is assigned on the basis of it's relative importance to other inputs.

- The node applies a function to the weighted sum of weights.

THE Idea is that the Synaptic stratergies (the weights- w) are learnable and can control the strength of influence and it's direction on one neuron on another

1. Excitory - (Positive Weight)

2. Inhibitory - (Negative Weight)

Biologically, 

                      The dendrite carry a signal to the cell body where they all get summed up. If a final sum is certain threshold, the neuron can fire, sending a spike along the Axon.

In computational model, we assume that the precise timings of the spikes do not matter, and that only the frequency of firing communicates information..

We model the firing rate of a neuron with the help of activation function . (Eg. Sigmoid function, Tanh function, Arctan function).

NEURAL NETWORK ARCHITECTURE

    When we train a neural network we want the neurons should fire once they learn some new specific patterns from the data. We model the fire rate using an activation function or(transfer function).

Let's just dig bit deeper into the actual parameter that we must know about a neural network:

•  INPUT NODES (INPUT LAYER) -  No computation is done here within this layer, they just pass the information to another nodes or to hidden layer or to output layer. The block of nodes is called layer.
• HIDDEN LAYER -  In the hidden layers intermediate computations are done, they perform computation and then transfer weights(Signal or information) from input layer to the subsequent layers( another hidden layer or output layer).
• OUTPUT NODES - Here we finally use an activation function to maps the output in desired format. (Eg. we use Softmax function for classification)
• CONNECTION AND WEIGHTS - The network consists of network of neurons. the output of one neuron is fed to another neuron as an input, their weights are also being connected while they are transferred from one node to another.
• ACTIVATION FUNCTION - An activation function of  a node defines the output of that node for the given input or set of inputs.
• It's the Non - linear function that allows to compute the non-trivial problems with such a network using small number of nodes. This function is also called as transfer function is artificial intelligence.
• LEARNING RULE - The learning rule of the neural network is to modify the parameters of the neural network, in order for a given input to the network to produce favored output.
• This learning process typically amounts to modify the Weights and Thresholds.

TYPES OF NEURAL NETWORK

There are many classes to a neural network and these classes can have many sub- classes, We will study the most used ones : 

1. Feed Forward Neural Network - This kind of neural network can be understood as the network which do not form a cycle or a loop.

- the information is forwarded in only one direction  - one node to another or to another layers.

They do not have something to learn back from as the neurons are only flowing the information is one direction the network is committed to defined Output for the defined input or set of inputs.

We  Have 2 Types Of Feed Forward Neural Network

1. SINGLE LAYER PERCEPTRON : It's the simplest feed forward network and does not contain any hidden layers, which means it only consists of input layer and the output layer.

- We do not count input layer as no computation are being performed on this layers and only information is collected and forwarded. 

- So, we only count Output layer which makes it a single layer perceptron.

   2. MULTI LAYER PERCEPTRON : This class of network consists of several nodes of computational units, usually interconnected in a feed forward manner.

- Each neuron is connected to form a layers and that is connected with subsequent layers.

- In many applications each units of these apply a Sigmoid function as an activation function.

-MLP's are very useful when it comes to non-linearity , they are able to learn non - linear representations(most of the cases the data represented to us is non linearity separable.

3. CONVOLUTIONAL NEURAL NETWORKS : CNN's - They are very similar to normal neural network.

- They have learnable weights and biases,

- In convolutional network the unit connecting pattern is inspired by the organization of visual cortex, units respond to stimuli - in a restricted region of space known as respective field.

- Respective fields partially overlap, over covering the entire visual field region.

- Units response can be can be approximated using mathematical convolution operation.

- They are variations of Multi Layer Perceptron that require minimal processing.

- Application Of CNN are wide : video recognition and Image recognition, Natural Language Processing.

- The only concerning factor with CNN's is that they require huge amount of data to train upon.

 RECURRENT NEURAL NETWORKS (RNN'S)

In recurrent neural networks, the main difference is there are connection between units that create a cycle or a loop( they propagate data to forward stages but also backwards, from later processing stages to early stages).

The feed back mechanism allows RNN's to exhibit dynamic temporal behavior.

- Unlike feed forward neural networks, RNN's got their internal memory to process arbitrary sequence of inputs.

This creates an immense possibility to work with RNN's- they are widely used with 

- Unsegmented , Connected handwriting recognition

- Speech recognition

- other general sequence processors.

If you found this article helpful, don't forget to subscribe to get more on these topics .

The next article will be including the most commonly used Activation Functions in Neural Networks.