Convolutional Neural Network: Guide to understanding them

This is the fundamental component of a CNN and where most of the calculations are performed. This layer uses a filter or kernel to move through the receptive field of an input image and detect the presence of specific features.

The process begins by sliding the kernel over the width and height of the image, and then traversing the entire image in several iterations. At each position, a scalar product is computed between the kernel weights and the image pixel values under the kernel. This transforms the input image into a set of feature maps, each of which represents the presence and intensity of a given feature at various points in the image.

Suppose the input is a color image, which is made up of a 3D array of pixels. Thus, the input will have three dimensions (height, width, and depth), plus a feature detector, which will move through the receptive fields of the image to check if the feature is present (convolution).

This detector or filter is a two-dimensional array of weights representing part of the image. Although it can vary in size, it is usually a 3×3 matrix, which also determines the size of the receptive field. This filter is applied to an area of the image and a scalar product is calculated between the input pixels and the filter. This scalar product is fed into an output matrix. The final result of the series of scalar products of the input and the filter is known as a feature map, activation map, or convolved feature.

Some parameters are adjusted during training through the process of backpropagation and gradient descent. However, there are three parameters that affect the size of the output volume that must be set before neural network training begins:

1. The number of filters affects the depth of the result.
2. Pitch is the distance or number of pixels the kernel travels over the input matrix. Although it is uncommon to find step values of two or more, a larger step produces a smaller result.
3. Zero padding is used when filters do not match the input image. This sets all elements outside the input matrix to zero, which produces a larger or equally sized output. The types are:
   1. Valid fill: the last convolution is skipped if the dimensions do not line up.
   2. Same fill: ensures that the output layer is the same size as the input layer.
   3. Full fill: increases the output size by adding zeros to the input border

After each operation, a CNN applies a rectified linear unit transformation to the feature map, introducing nonlinearity into the model.

Source: Developers Break

When a convolution layer follows the initial layer, the structure of the CNN can become hierarchical, as later layers can see pixels within the receptive fields of earlier layers.

Each individual part of the image forms a lower-level pattern in the neural network, and the combination of its parts represents a higher-level pattern, creating a hierarchy of features within the CNN. Finally, the convolutional layer converts the image into numerical values, allowing the neural network to interpret and extract relevant patterns.

Also known as downsampling, they perform dimensionality reduction, which reduces the number of parameters in the input.
Similar to the first layer, the clustering operation sweeps the entire input with a filter, but the difference is that this filter has no weights. Instead, the kernel applies an aggregation function to the values within the receptive field, filling the output matrix. There are two types:

• Maximum Clustering: as the filter moves through the input, it selects the pixel with the maximum value to send to the output matrix.
• Average clustering: as the filter moves through the input, it calculates the average value within the receptive field to send to the output matrix.

The disadvantage of this layer is that a lot of information can be lost, but it also helps to reduce complexity, improve efficiency, and limit the risk of overfitting.

The pixel values of the input image are not directly connected to the greeting layer in the partially connected layers. That’s where the fully connected layer comes in, where each node in the output layer is directly connected to a node in the previous layer.

This layer performs the classification task based on the features extracted through the previous layers and their different filters.

CNNs are designed for feature extraction by using convolutional layers to capture these hierarchical features from the input data. This is crucial for tasks such as recognition, segmentation, and object detection. In addition, they can learn discriminative features directly from the data.

GANs, on the other hand, are not intrinsically designed for feature extraction. Although their discriminator learns to differentiate features, it is not their primary purpose, as they focus on data generation and manipulation.

Source: LeewayHertz

As discussed earlier, the primary function of CNNs is to process existing data for classification, detection, or segmentation tasks, they are not typically used for data generation. However, there are some variants, such as variational autoencoders (VAE), that can be adapted for data generation.

On the contrary, GANs excel in this task, as they can create highly realistic images, making them the best choice for use cases such as image synthesis, super-resolution, and style transfer.

In this case, CNNs are well suited for transfer learning, as pre-trained models are available and easy to tune for specific tasks, which is very beneficial when working with limited data.

In turn, although GANs are less frequently used in this scenario, some models are pre-trained.

CNNs are discriminative models, as they focus on distinguishing and classifying data.

On the contrary, GANs are generative models, designed to create new data that can be distinguished from real data.

CNNs process existing images, not produce them, so their performance depends on the quality and quantity of data used for training. Although they are capable of achieving high accuracy in recognition tasks, they do not inherently generate realistic images.

GANs have the ability to produce images that can “fool” humans due to their realism, so they have set benchmarks in image generation and quality.

On the one hand, we find CNNs, which require a great computational effort during training, especially when working with deep architectures and large datasets. However, they are also relatively fast when making predictions once trained.

GANs are also computationally and training resource intensive. With these demands, real-time applications can be challenging.

Here we find similarities, as both face ethical and security challenges that have yet to be resolved. For example, CNNs have raised ethical concerns related to privacy, bias, and surveillance, especially with regard to facial recognition.

GANs are not far behind in these challenges, as their use to create deepfakes raises significant ethical and security issues, as the technology can be used for misinformation, impersonation, or other malicious purposes.

Therefore, both CNNs and GANs have their own sales and applications in the field of AI, especially in computer vision. The choice between one or the other depends on the specific requirements of the task at hand. CNNs are the ideal choice for tasks that involve recognizing and analyzing existing data, while GANs excel in tasks that require data generation, image synthesis, and creative expression.

However, these technologies are not mutually exclusive; in some applications, they can complement each other. As machine vision continues to evolve, both will play a key role in shaping the future of visual perception and understanding.