A Beginner ' s Guide to Understanding convolutional neural Networks Part One note

Original link: Https://adeshpande3.github.io/adeshpande3.github.io/A-Beginner ' s-guide-to-understanding-convolutional-neural-networks/

This article is a preliminary understanding of convolutional neural Networks (convolutional nerual Networks)

Image classification

Image classification ( images classified ) is the task of taking an input image and outputting a class (a dog, a cat, ect.) or a P Robablity of classes that's best describes the image.

Inputs and Outputs

When a computer sees an image, it would see an array of pixel values, e.g. 32*32*3, RGB (red,green,blue) values.

　　/**** Supplement ****/

Single channel diagram: Commonly known as grayscale, each pixel can only have one value for color, pixel value between 0-255 (0 is black, 255 is white, median is some different levels of gray).

Three-channel graph (RGB): Each pixel point has three values, the red, green, and blue channel values of the three color changes and the overlap between them to get a variety of colors. A three-channel grayscale image refers to the same value for three channels.

Biological Connection

Some neurons respond only to the edges in a particular direction, and some neurons respond only to the vertical direction, some to the horizontal direction, and so on. These neurons are in a columnar tissue (a light receptor in the human eye: a columnar body that has a general perception of things) and is the basis of a convolutional neural network.

First Layer-math part (convolutional layer aka conv layer)

　　

The filter (or a neuron neuron /kernel nucleus ) has a array of numbers,called weights or parameters. The filter is convolving, next step (stride) was moving to the right by 1 unit.

The depth of the same as the depth of the input, so the filter is 5*5*3. If We use the Filters (5*5*3), the output would is 28*28*2.

First Layer-high level perspective

Each of the these filters can be thought of as feature identifiers (straight edges, colors, curves ect.).

e.g. a curve detector

The filter'll has a pixel structure in which there'll be higher numerical values along the area that's a sh Ape of a curve.

　　

So we take this image as example.

　　

　　(Visible the first picture match degree is high, the second match degree is low)

Going deeper Through the Network

　　A Classic CNN Architecture would look like this:

ReLU, Conv, ReLU, ReLU, Conv, ReLU, pool, Fully, Conv, CTED Layer

(ReLU: Activation function, pool: pooling layer)

There ' re other layers that is interspersed ( embellishment, scatter ) between these conv layers, they provide nonlinearities (ReLU) and preservation ( Dimension protection ) of dimension (Pool) that help to improve the robustness ( robustness ) of the network and Co Ntrol overfitting.

as you go through + and more conv layers, (i). You get activation maps, the represent more and more complex features; (ii). The filters begin to have a larger and larger receptive field.

Fully Connected Layer (FC)

The whole connection layer plays the role of classifier in the whole network, and can be realized by convolution.

Currently full connectivity due to parameter redundancy (only the full connection layer parameter can account for the entire network parameter of about 80%), the recent use of global average pooling (average pooling,gap), usually has better predictive performance.

　　

A Beginner ' s Guide to Understanding convolutional neural Networks Part One note