Convolutional Neural Network (CNN) PDF

· CONVolutional NEURAL NETWORK. A Convolutional Neuval Network /CNN) is a specialized type of deep neural network to process structured data....

· CONVolutional NEURAL NETWORK. A Convolutional Neuval Network /CNN) is a specialized type of deep neural network to process structured data. unlike traditionally fully connected networks the , CNNs are equipped with layers that extract and process spatial or temporal information hierarchically : powerful for tasks that involve visual or sequential data. CONVOLUTIONAL LAYER. This layer applies a set of filters to the input data. Each filter slides the input dot product between over , computing a filter weights and input values. The result is a feature map that highlights specific features redges , textures, patterns ec. / RELU ACTIVATION. After the convolution operation a nonlinear activation function , Stypically the Recu) is applied element-wise. it to learn It introduces non linearity into the model , enabling complex pattern. So the curs are used widely in image recognition tasks where , , the goal is to assign a label to an image based on its content. Classic applications face recognition , object detection. CNNs architectures have achieved human-level performance in some tasks ; CNNs analyze X-rage. The field of natural networks employs diverse architectures. There are different categories that help to understand how different architectures operate · Supervised. Type of learning Unsupervised. Deep. Structural depth Shallow. 1 Supervised learning vs Unsupervised learning. Distinction based of the data and the objective on the type of the learning process. 4) Supervised. The model learns from the labeled data , meaning that for each input the output is provided , corresponding The goal is to map inputs to outputs accurately /minimizing differences between predicted and actual labels) Used in image classification object detection. , ecc. 2) Unsupervised Deals with unlabeled data. The goal is to uncover hidden patterns , structures or relationships in data without guidance on what the output should be. Used for clustering ecc. 2 Deep vo Shallow. Networks architectures are categorized as Deep or Shallow based on the number of layers in the model. 1) Deep architectures. Multiple layers hundreds enobling them /dozens or even to learn hierarchical and abstract representations of data. Particularly powerful for capturing complex patterns. 2) Shallow architectures. They typical have one or two layers. They are less capable of learning intricate patterns but they are , computationally efficient and often suitable for Simpler problems. So , we haveh different categories , one for each quadrant. 1) Supervised Deep + : Models that learn from labeled data and have many layers allowing, them to handle complex tasks Examples. : Convolutional Neural Networks /CNNS). Recurrent Neurol Networks (RNN). 2) Supervised + Shallow. Models that learn from labeled data but with fewer layers. Examples : Perceptions. Support Vector Machines (SVMS). Often preferred for small datasets tasks fast or requiring computation and high interpretability. 3) Un supervised + Deep. Architectures that aim to uncover patterns in unlabeled data through multiple layers Examples. : Deep Belief Networks (DBNS). sparse Denoising Autoencoders. Used for applications where discovering latent structures is Key. 4) Unsupervised + Shallow. Computationally efficient and widely used in exploratory data analysis Examples. : Restricted Boltzmann Machines (Rums). Autoencoders. Typically used when the dataset is relatively simple. BASIC COMPONENTS Of CNNs. CNNs are composed of a series of interconnected layers that work together to transform input data into meaningful patterns used for classification , detection or segmentation. layers extract increasingly complex and abstract features from the input data. There are three types of layers in a CNN : · The convolutional layers , cre building blocks , responsible for extracting features from input data using filters that slide across the image , detecting patterns such edges , textures e · The pooling layers , which downsample the feature maps to reduce spatial dimensions and computational complexity while preserving important features. · The Fully connected layers , that act as the decision-making extracted features into final components , transforming the a output , like classifications or predictions. More specifically : CONVOLUTIONAL Layers. Primary purpose is to opatial and hierarchical features extract like images by from in put data , performing convolution operations. , Small learnable filters are applied to the input data. POOLING Layers. Designed to reduce the spatial dimensions of feature maps while the most important informations. preserving By progressively reducing the size of feature maps , pooling layers help control overfitting and decrease computational complexity. This makes the model more robust and this property is called Shift invariance. They are typically placed between successive convolutional layers in a cNN architecture. Some operations can be done in pooling layers They. are performed local regions over called pooling windows. For each pooling window a specific operation is applied and then the result is stored in corresponding location of the output feature map. The most common are : 1) Max pooling. Selects the maximum value from each window pooling. Widely used because it captures the most prominent feature within a region. 2) Average pooling Computer the average value of all elements in the pooling Window. It is useful in certain applications where capturing the average response of a region is more meaningful thon selecting the maximum activation. FULLy Connected Layers. They serve as bridge between feature extraction process/ performed a by convolutional and pooling layers) and the output layer/which provides predictions). They estabilish connections between every neuron in the current layer and every neuron in the previous lager. Dense connectivity enables fully connected layers to all extracted features and integrate the perform high-level reasoning. y = f(wx + b). So their , primary role is to integrate and interpret features extracted by previous layers. They operate at the end of the CNN , where spatial dimensions of the feature maps are significantly reduced due to the pooling operations. They perform at high-level reasoning. Convolutional layers may recognize specific features of a dog reges tail , , ecc. ), the fully connected layer combine these feature to conclude that the image contains a dog. despite of this , they have very high computational cost and risk of a overfitting. To address these issues , 1x1 convolations have emerged. TRAINING Process. The loss function is a critical component of a cun , guiding the training process by quantifying the error between the networks prediction and the true labels Objective is to minimize this loss.. The training process in CNNS is a critical phase where the network learns to extract meaningful patterns and relationship from the input data to perform tasks such as classification , regression or object detection. Training involves iteratively updating network's parameters , so that the cun can minimize the difference between its predictions and the actual target values. this is achieved by the gradient descent.