The Fundamentals of Neural Network | CNN
What is a Neural Network and how it works? We are gonna cover everything!!!
Let’s begin this article with a video.
It feels amazing to take decisions on our own and trying new things. We also feel strong after learning from our mistakes. We still do not know much about how a human brain works. The brain could work by learning the strength of connections. We got to know about neurons in our brains and the way how it works. So it would be much better if the computer could learn by itself.
How a biological neuron works?
When a child sees a cat for the first time they do not know that it is a cat, until someone says so. But later after the child sees a lot of cats they are able to identify by themselves. It is due to the reason that the brain now been trained by a lot of information on how a cat looks and it compares that information to the object they see and defines that it is a cat. This event takes place in the huge neural network which is made up of neurons.
How do neurons work in machine learning?
A neural network is a system that mimics the way how the neurons in our brain work. This system consists of a series of algorithms that recognize the relationship in the set of data.
The neurons get trained by the inputs and this input forms certain connections or relationships that it is able to predict the output when a new input is given. The inputs get multiplied with the weights and added with the bias and the output is given. Weight is the strength of the connections between the neurons. The model decides by itself the bias value based on the threshold value. It improves the flexibility of the output.
A single neuron is a linear function without the activation function. The activation function is the deciding factor. It helps the neural network to learn complex patterns. This means that it helps the neural network to map non-linear functions as well.
Sigmoid, tanh, and RELU are few examples of the activation function.
Did you know? The neural network has a huge part in speeding up video games and making graphics more realistic.
Before jumping ahead let’s rewind to the year when the first neural net was created.
Frank Rosenblatt in 1957, was the first to introduce a neural network. He created a model known as Perceptron which classifies the images of men and women. It consisted of only a single layer of artificial neurons, due for this reason it didn’t work very well.
Later Geoffrey Hinton devised a model with multiple layers. These layers are known as hidden layers. Each hidden layer can have different activation functions. It was more productive than the perceptron model but the computers in those days were not powerful enough to handle it and there was not much data available.
Here let’s know a term called “Deep neural network”
Deep Neural Network
The Neural Network with a multilayer approach is known as a deep neural network.
There are layers of neuron which communicate in such a way that each artificial neuron learned a specific feature from the data. Each input will be connected to the hidden layer. So that it can predict the answer to the data it has never seen before.
As a result in the late ’80s, self-driven cars were developed. Later in the ’90s Yann LeCun developed a model that can detect handwritten numbers, which gave birth to a new type of network known as the Convolutional Neural Network or the CNNs.
Convolutional Neural Network
Convolutional Neural Network is an artificial neural network. It is popularly used for analyzing images.
The convolutional layers in hidden layers detect the patterns in the image like edges, corners, and geometric shapes like circle, square, triangle. These are done by the filters. The deeper the network goes the more sophisticated the filters become. In later layers, the filters would be able to detect specific objects like eyes, ears, hair, and so on.
We have to specify the number of filters the layer can have while adding the convolutional layers. Filters are the small matrix for which the values are randomized. Here let’s assume a 3 by 3 matrix. When this convolutional layer receives in just the weights in the simple neural network, the filter will convolutional operation to the pixels from the input until it covers the entire image. Thus the image is detected at the end.
Application of Convolution Neural Nets
- Neural Style Transfer
- Tesla Auto-pilot | Self-Driving Car
- X-ray detection
- Cancer Cell detection: Check the link to the Kaggle notebook
- Categorizing Listing Photos at Airbnb
- Facebook research on CNN | Detectron
When a machine learns on its own rather than being pre-programmed, it makes it work faster and more efficiently. By doing so, the computer can interact and process real-time data.
Further Recommendations:
- Check out this playlist on Neural Nets by Grant Sanderson.
- The Candian Genius.
- CNN by Stanford University.
- Neural Network interactive visualization by Tensorflow by Google.
Previous Article: Linear Regression | The idea of the old and the model of the future
