How Convolutional Neural Networks Work

Neural networks are a subfield of artificial intelligence and are at the core of deep learning models. They are made up of node levels, each of which includes an input layer, an output layer, and one or many hidden layers. Every node does have a threshold and weight that are connected to one another. Every node whose output exceeds the defined threshold level is triggered and begins providing data to the network’s uppermost layer. Or else, no data is transmitted to the network’s next tier.

There are other kinds of neural network models, which are utilized for diverse use instances and data types. Recurrent neural networks, for instance, are frequently used for speech and natural language processing, but convolutional neural networks (also known as ConvNets or CNNs) are more frequently employed for computer vision and classification applications. Before CNN's, identifying objects in images required the use of laborious, manual feature extraction techniques. Convolutional neural networks, on the other hand, now offer a more scalable method for classifying images and recognizing objects by using matrix multiplication and other concepts from linear algebra to find patterns in images. However, they can be computationally taxing, necessitating the use of graphics processing units (GPUs) when modeling them.

How Convolutional Neural Networks Work

Convolutional neural networks outperform other neural networks when given inputs such as images, voice, or audio, for example. There are three basic categories of layers in them:

● Fully-connected (FC) layer
● Convolutional layer
● Pooling layer

The convolutional layer is the first layer of a convolutional network. The fully-connected layer is the final layer, though additional convolutional or pooling layers may be added following convolutional layers. With each additional layer, the CNN grows more intricate and is able to recognize greater portions of the image. Early layers emphasize fundamental components like colors and boundaries. The target object is eventually identified as the visual data progresses through the CNN layers and starts to discern larger components or properties of the object.

Convolutional Layer

The central component of a CNN is the convolutional layer, which is also where the bulk of computation takes place. It needs a filter, input data, and a feature map, among other things. Suppose that the input will be a color picture that is composed of a 3D array of elements. As a result, the input will also have 3 dimensions—width, depth, and height,—that are analogous to RGB in a picture. Additionally, we have a feature detector, also referred to as a  filter or kernel, which will traverse through the image’s receptive fields and determine whether the feature is there. Convolution describes this method.

A 2-dimensional array of weights serving as the feature detector reflects a section of the image. The filter size, which also controls the dimension of the receptive field, is normally a three-by-three matrix, however, they can differ in size. Following the application of the filter to a section of the image, the dot product between both the filter and the input pixels is determined. The output arrays is then fed with this dot product. Once the kernel has swept through the image sequence, the filter adjusts by a stride and repeats the operation. An activation map, feature map, or convolved feature is the ultimate result of the sequence of dot products from the filter and the input.

Every output variable in the feature map does not necessarily have to correspond to every pixel value in the target image. Only the receptive field, in which the filter is being used, must be connected. Convolutional (and pooling) layers are repeatedly alluded to as “partially connected” layers because not every variable in the output array has to map exactly to every variable in the input array. The local connection is another name for this property.

It should be noted that the feature detector, sometimes referred to as parameter sharing, keeps its weights constant as it advances over the picture. Through the processes of gradient descent and backpropagation, various variables, such as the weight values, modify during training. However, before the neural network training process starts, 3 hyperparameters that determine the output volume size must be established. These consist of:

The depth of the output is affected by the number of filters. Three major filters, for instance, would produce three various feature maps, giving a depth of 3.

The kernel’s stride is the number of pixels it travels over the input matrix. Despite the rarity of stride values of two or higher, a longer stride results in a lesser output.

In cases where the filters don’t always accommodate the input image, zero-padding is typically utilized. This results in a larger or similarly sized output by setting any elements that are not part of the input matrix to zero.

A CNN performs a Rectified Linear Unit (ReLU) adjustment on the feature map following each convolution function, adding nonlinearity to the models.

Pooling Layer

Downsampling, sometimes referred to as pooling layers, carries out dimensionality reduction and lowers the number of variables in the input. The pooling layer runs a filter throughout the entire inputs similarly to the convolutional layer, with the exception that this filter lacks weights. Rather, the kernel populates the output array by applying an aggregation function to the variables in the receptive field.

The pooling layer eliminates a lot of information, but it also provides CNN with several benefits. They reduce complexity while increasing efficacy and decreasing the chance of overfitting.

Fully-Connected Layer

The full-connected layer is exactly what its name implies. As was already noted, partially connected layers do not have a direct connection between the output layer and the input picture’s pixel values. In contrast, every node in the output layer of the fully-connected layer is intrinsically linked to a node in the layer above it.

Depending on the features that were retrieved from the preceding layers and their various filters, this layer conducts the classification operation. FC layers often utilize a softmax activation function to categorize inputs adequately, generating a probability ranging from 0 to 1. Pooling and Convolutional layers typically use ReLu functions.