What do you understand by the term CNN (Convolutional neural networks)????

Convolutional neural networks(CNN) is a class of Deep neural networks, most commonly used to analyze the Image data and pattern recognition. It can take in an input image, assign importance (labels and weights) to various objects in the image and it is able to differentiate one from another. The pre-processing time required for CNN's are much lower when compared to other classification algorithms.

The picture shows the model of a CNN. The input is the image of size 28*28*1 which refers to width*height*depth of image. If the image is gray scale then we only have one color channel, so the depth of the gray scale image is one. If the image is colored image then it has three color channels (depth is 3) namely red, blue, green. Each color channel is individually of two dimensional and together it is of three dimensional. A filter (kernel) is applied on the image and it is converted into a convoluted image. we can specify the padding after filter is applied. The convoluted image is then converted into new image by pooling the convoluted image. We can specify these iterations according to our needs. Then these three dimensional images are flattened and passed through dense layers for classification. We can specify the optimizer, loss, layers, padding, kernels, strides and activation functions in the network. This is the basic model of convolutional neural networks. 

Word Description        

Layer: A collection of nodes connected together within a neural network.

Dense layer: A fully connected layer. Each node in one layer is connected to each node in previous layer.

Optimizer: Gradient descent is an algorithm that changes the internal weights a bit at a time to gradually reduce the loss function. Optimizer is a specific implementation of gradient descent algorithm. There are many optimizers available, the most commonly used is 'adam'. 

Flatten: The process of converting 2D or 3D image into vector.

Convolution: The process of applying filter(kernel) to an image.

Kernel: A matrix smaller than input image used to transform the input image into chunks.

Padding: Adding the pixels into image. Normally we use zero padding.

Pooling: The process of reducing the size of an image through down sampling. There are several types of pooling layers. Average pooling layer converts many values into a single value by taking average. However, max pooling is the most commonly used pooling layer. 

Strides: Number of pixels to slide the kernel across the image.