Convolutional Neural Networks Basics

The number of parameters in a neural network layer, calculated by multiplying the number of units in the layer by the number of inputs to each unit.

A fully connected neural network (FCN) treats an image as a flat vector of pixels, ignoring the spatial relationships between pixels. This leads to the loss of spatial correlation, which can be detrimental for some image recognition problems.

Features in an image, such as edges, corners, and endpoints, are key elements for recognition. These features are spatially correlated and provide information about the image's structure.

A technique for automatically learning feature detectors within a convolutional neural network (CNN), where the network itself discovers the most relevant features to recognize objects in images.

A type of neural network that is specifically designed to process spatial data, such as images. CNNs use convolutional operations to extract features from images, allowing them to recognize patterns and objects.

A mathematical operation that involves sliding a filter (kernel) across an input signal (e.g., an image), computing the dot product between the filter and the local region of the input. This process helps identify patterns and features within the input.

A convolutional neuron is a type of artificial neuron that performs a 2D convolution operation on an input image.

Feature extraction using convolution is a process of extracting feature maps from an image by performing convolution operations with different filters.

A feature map is a representation of the extracted features from an image using a filter.

A 3x3 feature map is often used and refers to a feature map with a size of 3 rows by 3 columns.

A filter is a small array of weights that are multiplied with the corresponding pixels in the image during convolution.

The weights in a filter are determined during the training process of a neural network and they learn to represent specific features in the input data.

The size of a filter can vary, but a common size is 3x3, which refers to a filter with 3 rows and 3 columns.

The output of the first convolutional layer, which contains the initial feature maps, is then fed into the second convolutional layer. This process continues for subsequent layers, with each layer extracting increasingly complex and abstract features from the image.

The output of a convolutional layer is often passed through a pooling layer, reducing the spatial resolution of the feature maps. This downsampling helps to simplify the information and make the model more efficient.

The process of increasing the number of feature maps in subsequent convolutional layers, allowing the model to extract more complex and detailed features from the image. This helps to improve model performance.

The technique of using smaller, more efficient convolutional layers instead of large fully connected layers in a neural network.

The action of reducing the size of the feature map, typically by taking the maximum value within a filter.

A layer of a convolutional neural network responsible for transforming input data through a series of specialized filters. It extracts features from the input data, highlighting patterns and important information.

A technique used for reducing the number of parameters in a convolutional neural network by using layers that learn a smaller set of features, leading to a leaner and more efficient model.

The spatial relationship between pixels in an image.

A revolutionary architecture in deep learning that utilizes residual connections to allow for much deeper networks, improving performance and reducing overfitting.

The ability to train significantly deeper networks by introducing residual connections, allowing for information to flow more easily even through many layers.

A key innovation in ResNet that allows information to bypass certain layers, enabling gradient flow and preventing vanishing gradients in deeper networks.

An image dataset used for benchmarking image recognition models, containing millions of labeled images.

A metric used to evaluate image classification models, representing the percentage of correctly classified images in the top 5 predictions.

A small matrix of weights that is applied to a local area of the input image to create a feature map in a convolutional neuron. The filters can be learned via backpropagation, allowing the network to automatically identify relevant features without manual engineering.

The process of applying a learnable filter to an input image to generate a feature map. This process is done by sliding the filter across the image, computing the weighted sum of pixel values at each location, and producing an output value.

A technique used to prevent the loss of information at the edges of the input image during the convolution process. It involves adding a border of zero-valued pixels around the image, ensuring that the convolutional filter can operate on all pixels without cutting off the edges.

In a convolutional neural network, multiple input channels represent different color channels of an image, e.g., red, green, blue (RGB). Each input channel corresponds to different information in the image, similar to how our eyes perceive different wavelengths of light.

In a convolutional neural network, a separate filter is applied to each input channel. This allows the network to extract different features from each channel, leading to a more comprehensive understanding of the image.

In convolutional neural networks, features detected from different input channels are combined together, often by summing the results of convolutions. This allows the network to learn more complex features by integrating information from different color channels, leading to a more powerful representation of the image.