CONVOLUTIONAL NEURAL NETWORKS
- Genaro Pimienta
- 2 days ago
- 2 min read
CONVOLUTIONAL NEURAL NETWORKS
Convolutional neural networks (CNN) are a type of deep learning architecture designed to extract features from image inputs. CNNs are commonly used in computer vision and image processing applications.
CNNs have a three-layer architecture (Figure 1).
Convolutional layer — feature extraction
Pooling layer — dimension reduction
Feedforward network — output classification
Figure 1. Basic CNN structure architecture. Features from an input (e.g., an image) are extracted with multiple convoluting filters, each of which populates a convolutional layer. After simplification and flattening the information in the convolutional layers enters a feedforward network for feature classification.
Compared to prototypical feedforward neural networks, CNNs extract features from numerical matrices, which can be two- or three-dimensional.
Feature extraction from an input is performed by multiple convolving filters (Figure 2).
Convolution filters — In CNNs, input features are extracted with a convolution filter. Convolution filters carry learnable weights, used to activate artificial neurons in an activation map; and have grid-like dimensions that cover a region in the input matrix, named receptive field. So, if a convolution filter has a 3x3 grid size, it means that it covers nine values in the receptive field.
Feature extraction — During feature extraction, the convolution filter “convolves” across the input matrix, advancing by strides of a few grid values at a time. At each position, the input values in the receptive field are multiplied by the weights in the convolution filter and summed up for neuron activation.
Activation maps — CNNs learn hierarchically by using different filters each trained to extract a specific type of content in the input. Since there is one filter per activation map, a convolutional layer is often composed of multiple activation maps.
Figure 2. Feature extraction in CNNs.
In shotgun proteomics, CNNs are used to predict fragment ion masses and retention times. Relevant examples are MS2CNN for fragment ion prediction and DeepLC for retention time prediction.
Comments