Convolutional Neural Networks Programming Strategic consultants in Data Science. CNN is an artificial neural network where “neurons” correspond similarly to receptive fields. This type of network is a difference of a multilayer perceptron. However, since their application is carried out in two-dimensional matrices, they are very effective for artificial vision tasks. Such as image classification and segmentation, among other applications.

Similarities With The Human Brain Convolutional Neural Networks Programming

Work done by Hubel and Wiesel in 1959 played an essential role in understanding how the visual cortex functions, particularly the cells responsible for orientation selectivity and edge detection on visual stimuli within the V1 primary visual cortex. Two cell types were identified because they have elongated receptive fields, making them more responsive to elongated visual stimuli such as lines and edges. These are called simple cells and complex cells. Moreover, Simple cells have excitatory and inhibitory regions, forming elongated elemental patterns in each cell’s particular direction, position, and size. Suppose a visual stimulus arrives at the cell with the same orientation and position so that the cell aligns perfectly with the patterns created by the excitatory regions and simultaneously avoids activating the inhibitory areas. In that case, the cell is activated and emits a signal.

Complex cells operate similarly. Like simple cells, they have a particular orientation to which they are sensitive. However, they are not position sensitive. Therefore, a visual stimulus must arrive only in the correct direction to activate this cell.

How Are They Built, And How Do They Work Convolutional Neural Networks Programming

 

Convolutional neural networks consist of multiple layers of convolutional filters of one or more dimensions. After each layer, a function is usually add to perform the non-linear causal mapping. Furthermore, like any network used for classification, in the beginning, these networks have a feature extraction phase composed of convolutional neurons. Then, there is a reduction by sampling, and at the end, we will have more superficial perceptron neurons to perform the final classification of the extracted features. In addition, The feature extraction phase resembles the stimulating process in the cells of the visual cortex. Moreover, this phase is composed of alternating layers of convolutional neurons and down sampling neurons. As the data progresses through this phase, its dimensional decreases, with neurons in distant layers. Being much less sensitive to disturbances in the input data but at the same time being activated by increasingly complex features.

How Do You Make A Convolutional Neural Network Programming Learn

Convolutional neural networks, CNN, learn to recognize diverse objects within images. Still, they need to “train” in advance with a significant number of “samples” -read more than 10,000. However, In this way, the neurons of the network will Be able to capture the unique characteristics -of each object- and, in turn, generalize it this is known as the process of “learning an algorithm.” So, for example, in this first convolution, we could have 32 filters, with which we will obtain 32 output matrices (this set is known as “feature mapping”).

Lastly,  Each one of 28x28x1 gives a total of 25,088 neurons for our FIRST HIDDEN LAYER of neurons, and that only analyzes a square image of just 28 pixels.

Comparison Between A “Traditional” Neural Network And A Convolutional Neural Network Programming

This table summarizes the differences between Fully connected networks and Convolutional Neural Networks.

FEATURES	“Traditional” Multilayer Feed forward Network	CNN Convolutional Neural Network
Input data in the Initial Layer	The characteristics that we analyze. For example, width, height, thickness, etc.	Pixels of an image If it is color, it will be three layers red, green, blue
hidden layers	we choose several neurons for the hidden layers.	We have the following types: * Convolution (with kernel size and some filters) * Subsampling
Output Layer	The number of neurons that we want to classify. To “buy” or “rent” will be two neurons.	We must “flatten” the last convolution with one (or more) layers of “traditional” hidden neurons and output using SoftMax to the output layer that classifies “dog” and “cat” will be two neurons.
Learning	supervised	supervised
interconnections	Between layers, all neurons move from one layer to the next.	There are fewer necessary connections because the weights we adjust will be those of the filters/kernels we use.
Meaning of the number of hidden layers	It is something unknown and does not represent something in itself.	The hidden layers are image feature detection maps and have a hierarchy: first layers detect lines, then curves and increasingly elaborate shapes.
Back propagation	It is uses to adjust the weights of all the interconnections of the layers	It is uses to adjust the consequences of the kernels.

Basic architecture (ABSTRACT)

• Input:   These will be the pixels of the image. They will be height, width, and depth will be one only color or 3 for example, Red, Green, and Blue.
• Convolution Layer:  will process the output of neurons that are connect in input “local regions” (i.e., nearby pixels). calculating the dot product between their weights (pixel value) and a small region they are connected to in the volume input. Here we will use, for example, 32 filters or the amount that we decide, which will be the output volume.
• RELU layer ; Moreover, will apply the activation function on the matrix elements.
• or SUBSAMPLING:  It will reduce the height and width dimensions, but the depth is maintain.
• “TRADITIONAL” LAYER:  feed forward network of neurons that will connect with the last subsampling layer and end with the number of neurons we want to classify.

 Conclusion

Convolutional Neural Networks As the data progresses through this phase. Its dimensionality decreases, with neurons in distant layers being much less sensitive to disturbances in the input data but at the same time being activated by increasingly complex features. Moreover, In this way, the neurons of the network will Be able to capture the unique characteristics -of each object- and, in turn, generalize it – this is known as the process of “learning an algorithm.” So, for example, in this first convolution, we could have 32 filters, with which we will obtain 32 output matrices (this set is known as “feature mapping”).