Where does CNN fit in Artificial Intelligence Puzzle?

Artificial Intelligence aims to bridge the gap between machines and humans by bringing about intelligent behavior in machines. Machine learning is one of the sub-sets of AI, that helps to bridge this gap. In Machine Learning we expect the machines to learn on their own by providing data, and without explicit programming. ML has sub categories such as Supervised Learning, Unsupervised Learning and Reinforcement Learning. The ML models are based on the data provided by Humans; hence all the flaws, biases and prejudices of human reasoning get very easily picked up by the model. In ML the most important problem is Feature Engineering which deals with features to be extracted to feed into the model. If the extracted features are incomplete then the model becomes biased and if all the features are extracted then again it causes overfitting, hence the dataset has to be too large if we are considering many features. Hence Neural Networks or Artificial Neural Networks comes into picture.

Artificial Neural Networks are considered as the foundation of AI. These networks are inspired by biological structure of a neuron and are capable of both, machine learning and pattern recognition. A neural network is based on the concept of a perceptron also known as the artificial neuron. A perceptron includes an input, multiplied with weights and a bias is added for activation. A layer of these perceptron, forming a dense connection called layers. Typically, a basic neural network consists of three layers called the Input Layer, the Hidden layer and the Output layer. The input layer is where we give the input information, and the output is some sort of prediction or classification. When there is more than one layer in the hidden layer then it is called a Multilayer Perceptron. Since the information is passed from one layer to another it is also called a Feed Forward Neural Network. The initial layers of the neurons extract the obvious features, but as the number of layers increase, non-trivial, subtle, and unique features are extracted. This is what makes the Artificial Neural Network “Deep”. There are different types of Deep Artificial Neural Networks based on the architecture and the computation, such as the Multilayer perceptron which uses Back Propagation Algorithm, Convolutional Neural Network which uses Convolution Operation and the Recurrent Neural Networks.

The term Deep Learning is used when we refer to a neural network which has more than three layers. In many ways DL is considered superior to ML because it can deal with a much larger dataset and the prediction mechanism is self-administered in DL whereas in ML, the programmer has to check the output and make changes accordingly.

When it comes to Deep Learning, CNN and MLP are often compared with each other. A basic MLP consists of three layers, namely, the input layer, the hidden layer and the output layer. Input is fed at the input layer and output is fetched at the output layer. We can have as many no. of hidden layers as we want. In an MLP every node/neuron of current layer is connected to every other neuron in the next layer. Each of the neurons have weights and biases. These parameters are all trainable i.e. in MLP all the parameters are trainable. We pass the input to model and multiply with weights and add bias at every layer and find the calculated output of the model. In MLP we use the Back-Propagation Algorithm. The loss is calculated and we back propagate the loss. According to which the weights are updated/altered.

Convolution Neural Networks or ConvNet or CNN is one of the important algorithms of AI which is mainly used in Computer Vision. CV aims to make the world interpretable to machines just like humans. CNN mainly take the input in the form of an image. The output is classification of some sort and the pre-processing required in CNN is very less compared to any other algorithm. In CNN we can observe a distinct feature; the number of parameters that are to be optimized are reduced and the weights are reusable. These networks can be trained even for a sophisticated image.