Introduction to Deep Learning.



Original Source Here

Introduction to Deep Learning.

Deep Learning is a fascinating subject and a very popular one too. Once you start understanding the concepts and their applications, you will realize the depth of the field and why it is so popular! DL is gonna change the world. It can be implemented in a lot of fields like medical, business, and many more.

DL is a complicated subject. There is a lot of math behind it. We are not gonna focus on too much math today. We are only talking about the introductory part. We will address the common questions asked about DL.

So, What is Deep learning anyway?

Deep learning is a class of AI that extracts higher-level features from raw input. Input can be matrices, images, audio files, etc.

Why Deep learning? (ML vs DL)

Machine learning has a lot of applications but it has few limitations. In ML feature extraction is a challenge, especially with large data. Few
applications like object recognition, handwriting identification, speech recognition are not possible. Deep learning is used to deal with this problem. DL models can extract the right features on their own and deal with large datasets as well.

How do Deep Learning models work?

DL uses neural networks. They are called artificial neurons aka perceptrons. They mimic the working of biological neurons. DL models take the input provided, runs these input data through several layers to predict and classify them. In DL, learning can be supervised, semi-supervised or unsupervised.

photo credits: towards data science

Now let’s compare biological neurons and perceptrons. The dendrites in the biological neuron receive an electrical impulse, which is processed inside the cell body then moved along the axon, when a specific synapse is reached it releases another impulse and passed to the next neuron. Similarly, in artificial neurons inputs are gathered (each input has a weight of its own representing each input’s significance), passed through net sum, activation function along with bias, and then output is obtained.

What is activation function?

credits: analytics vidya

The activation function is used to define the output for a given set of inputs. Types of activation function are sign, sigmoid, linear, and ReLu.

What is Artificial Neural Network (ANN)?

credits: tech Vidvan

ANN is the basic neural network. The working of this network is very simple. It works the same as the perceptron but it consists of multiple hidden layers between input and the output.

What are the types of neural networks??

There are mainly three types of neural networks. Well, there are more. These are the main types.

  1. Feedforward Neural Network.
  2. Recurrent Neural Network (RNN)
  3. Convolutional Neural Network (CNN)

Now, We will learn briefly about these three.

What is Feedforward Neural Network?

In this neural network, the information always flows in one direction (hence the name feedforward). These are sometimes also called multilayer perceptrons. As the flow is unidirectional there is no feedback of the information from the output. For an input x, this network provides us an output y. Inputs can be in the form of scalar, vector, matrices, or even an image. There are multiple hidden layers and each layer has a specific function. Through each layer, a function f(x) is performed. Mathematically, the formula for n layers can be given as

credits: the University of Toronto, Coursera

where, f(n)=output layer and f(1) to f(n-1)=hidden layers.

Limitations:

  1. Loss of previous/neighbor information.
  2. Its architecture only accepts and produces a fixed-size vector.

What is Recurrent Neural Network (RNN)?

Recurrent Neural Networks solve the problem in feedforward networks. RNN uses a backpropagation algorithm, in which the output of the iteration is fed back as input to the next iteration. Let us understand this with an example. Say there 7 chores spread over a week in a particular sequence. Each chore depends on whether or not the previous chore was completed. Like 2 comes after 1, 3 after 2, similarly till 7. To make this possible, we need to have information on the previous chore. This is where RNN comes into the picture as it is not possible in feedforward networks.

credits: Morioh

What is Convolutional Neural Network (CNN)?

credits: UpGrad

CNN is also like a feedforward neural network. It is inspired by the visual cortex of the human brain. CNN is mainly used in image classification (or any applications involving images). CNN has multiple convolutional layers which help in detecting the patterns in an image. These patterns are then used as features to classify and predict the output. CNN reads this image as a matrix. CNN has four steps,

  1. Convolution
  2. ReLu layer
  3. Pooling
  4. Fully connected

As this is just an introductory part, we are not gonna go into the details of each step.

Written by, Smruthi R Paladhi.

AI/ML

Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot

%d bloggers like this: