Artificial intelligence is the branch of data science which is based on the concept of human intelligence where machines demonstrate their intelligence in the similar way as human does. This term is applicable to any machine which is capable of learning and performing intelligent tasks. This field finds vast application in various fields such as image processing, speech recognition, etc.
Why Neural Networks?
Computers cannot solve the problems unless we specify what task it must do. We must provide the instructions to the computer as to what to do, how to do and when to do. The problems whose solution is not known; how do we deal with those? With Neural networks, the computers can perform tasks which we wouldn’t know how to solve them. Neural networks have proven to be revolutionary in the field of data science.
Neural Networks learn from an example, the data is provided to the network which acts as training set. We cannot program neural networks to perform any specific task.
The building block for neural networks is neuron. Like the human biological brain, the artificial neural networks functions. Such systems have capability to learn and progressively improve the performance to do task based on the learnt samples without the task specified program. This is the reason artificial intelligence is being focused on.
How it works?
Single neuron is called as Perceptron. It takes input which is the number of features and based on certain activation function, the output is generated.
The inputs will have weights assigned to them and the sum of all inputs will trigger the neuron and the output is generated. There is an activation function which is used as trigger in the perceptron, which controls the value of the output.
Modes in Neural networks
The input is fed to the neurons, the neuron identifies the input pattern and based on the pattern learnt, it decides whether to trigger the neuron or not. Thus, more the quality input, better the learning. This is training mode as we are training our model to learn from the input.
Once the model is trained, it is all set to work on the unseen data. The data which the model has not already seen is the unseen data. The model is fit on this data and as per the trained data; model will assign the weights to these input observations and calculates the predicted outcome.
Weights play very important role when building our neural networks. The features which are very important will have high weights assigned to it, while the least important features will have the least weights.
The weights assigned initially will have some random values for all the inputs. The output is calculated with these weights, the predicted output and the actual output are compared to get the error. The error calculated will be reduced by adjusting the weights at every iteration. The weights are adjusted based on the concept to get the global minima for the error.
The single perceptron will not be able to perform complicated tasks and hence we need something more than that. We combine multiple perceptron’s which are connected to each other to form a single network. The connection can be in many ways and may use different activation functions within a single layer (Layer refers to group of perceptron’s present at the same level).
Looking at the way the world is evolving there is no doubt that artificial intelligence will be the majority in most of the sectors.