CNN vs RNN: Difference Between CNN and RNN

Introduction 

In the field of Artificial Intelligence, Neural Networks which are inspired by the human brain are widely being used in extracting and processing complex information from various data and the use of both Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) in such applications are proving to be useful.

In this article, we shall understand the concepts behind both the Convolutional Neural Networks and the Recurrent Neural Networks, see their applications and distinguish the differences between both the popular types of Neural Networks.

Learn Machine Learning Training from the World’s top Universities. Earn Masters, Executive PGP, or Advanced Certificate Programs to fast-track your career.

Neural Networks and Deep Learning 

Before we get into the concepts of both Convolutional Neural Networks and Recurrent Neural Networks, let us understand the concepts behind Neural Networks and how it is linked with Deep Learning.

In recent time, Deep Learning is once concept that is widely used in many fields and hence it is a hot topic these days. But what is the reason behind it being so widely spoken of? To answer this question, we shall learn about the concept of Neural Networks.

In short, Neural Networks are the backbone of Deep Learning. They are a set number of layers consisting of highly interconnected elements known as neurons which perform a series of transformations on the data which generates its own understanding of that data which we refer to the term, features.

What are Neural Networks?

The first concept that we need to get through with is that of Neural Networks. We know that the Human Brain is one of the complex structures to have ever been studied. Owing to its complexity there has been a huge difficulty in unravelling its inner workings but in the present, several kinds of research are being undertaken in revealing its secrets. This Human Brain serves as the inspiration behind the Neural Network models.

By definition, Neural Networks are the functional units of Deep Learning which utilizes these Neural Networks to mimic brain activity and solve complex problems. When input data is fed to the Neural Network, it is processed through the layers of perceptron and finally gives the output.

A Neural Network consists of basically 3 layers – 

• Input Layer
• Hidden Layers
• Output Layer

The Input Layer reads the input data which are fed into the Neural Network system for further pre-processing by the subsequent layers of artificial neurons. All the layers that exist between the Input Layer and the Output Layer are termed as the Hidden Layers.

FYI: Free Deep Learning Course!

It is in these Hidden Layers where the neurons present in them make use of weighted inputs and biases and produces an output utilizing the activation functions. The Output Layer is the last layer of neurons that gives us the output for the given program.

Source 

How Do Neural Networks Work?

Now that we have an idea of the basic structure of Neural Networks, we shall go forward and understand how they work. To understand its working, we have to first learn about one of the basic structures of Neural Networks, known as the Perceptron.

Perceptron is a type of Neural Network which is the most basic in form. It is a simple feed-forward artificial neural network with only one hidden layer. In the Perceptron network, each neuron is connected to every other neuron in the forward direction.

The connections between these neurons are weighted because of which the information that is transferred between the two neurons is strengthened or attenuated by these weights. In the training process of the Neural Networks, it is these weights that are adjusted to get the correct value.

The Perceptron makes use of a binary classifier function in which is maps a vector of variables that are binary in nature to a single binary output. This can also be used in Supervised Learning. The steps in the Perceptron Learning Algorithm are – 

1. Multiply all the inputs with their weights w, where w are real numbers that can be initially fixed or randomized.
2. Add the product together to obtain the weighted sum, ∑ wj xj
3. Once the weighted sum of inputs is obtained, the Activation Function is applied to determine whether the weighted sum is greater than a particular threshold value or not depending upon the activation function applied. The output is assigned as 1 or 0 depending upon the threshold condition. Here the value “-threshold” also refers to the term bias, b.

In this way, the Perceptron Learning algorithm can be used to fire up (value =1) the neurons present in the Neural Networks that are designed and developed today. Another representation of the Perceptron Learning Algorithm is – 

f(x) = 1, if ∑ wj xj + b ≥ 0

0, if ∑ wj xj + b < 0

Though the Perceptrons are not widely used nowadays, it still remains as one of the core concepts in Neural Networks. On further research, it was understood that small changes in either the weights or bias in even one perceptron could vastly change the output from 1 to 0 or vice versa. This was one major disadvantage of the Perceptron. Hence, more complex activation functions such as the ReLU, Sigmoid functions were developed which introduces only moderate changes in the weights and bias of the artificial neurons.

Source 

Convolutional Neural Networks  

A Convolutional Neural Network is a Deep Learning Algorithm that takes an image as an input, assigns various weights and biases to various parts of the image such that they are differentiable from each other. Once they become differentiable, using various activation functions the Convolutional Neural Network Model can perform several tasks in the Image Processing domain including Image Recognition, Image Classification, Object and Face Detection, etc. 

The fundamental of a Convolutional Neural Network Model is that it receives an input image. The input image can be either labelled (such as cat, dog, lion, etc.) or unlabelled. Depending upon this, the Deep Learning algorithms are classified into two types namely the Supervised Algorithms where the images are labelled and the Unsupervised Algorithms where the images are not given any particular label.

To the computer machine, the input image is seen as an array of pixels, more often in the form of a matrix. Images are mostly of the form h x w x d (Where h = Height, w = Width, d = Dimension). For example, an image of size 16 x 16 x 3 matrix array denotes an RGB Image (3 stands for the RGB values). On the other hand, an image of 14 x 14 x 1 matrix array represents a grayscale image.

Source

Layers of Convolutional Neural Network  

As shown in the above basic Architecture of a Convolutional Neural Network, a CNN Model consists of several layers through which the input images undergo pre-processing to get the output. Basically, these layers are differentiated into two parts – 

• The first three layers including the Input Layer, Convolution Layer and the Pooling layer which acts as the feature extraction tool to derive the base level features from the images fed into the model.
• The final Fully Connected Layer and the Output Layer makes use of the output of the feature extraction layers and predicts a class for the image depending upon the features extracted.

The first layer is the Input Layer where the image is fed into the Convolutional Neural Network Model in the form of an array of matrix i.e., 32 x 32 x 3, where 3 denotes that the image is an RGB image with an equal height and width of 32 pixels. Then, these input images pass through the Convolutional Layer where the mathematical operation of Convolution is performed.

The input image is convolved with another square matrix known as the kernel or filter. By sliding the kernel one by one over the pixels of the input image, we obtain the output image known as the feature map which provides information about the base level features of the image such as edges and lines.

Convolutional Layer is followed by the Pooling layer whose aim is to reduce the size of the feature map to reduce computational cost. This is done by several types of pooling such as Max Pooling, Average Pooling and Sum Pooling.

The Fully Connected (FC) Layer is the penultimate layer of the Convolutional Neural Network Model where the layers are flattened and fed to the FC layer. Here, by using activation functions such as the Sigmoid, ReLU and tanH functions, the label prediction takes place and is given out in the final Output Layer.

Where the CNNs Fall Short 

With so many useful applications of the Convolutional Neural Network in visual image data, the CNNs have a small disadvantage in that they do not work well with a sequence of images (videos) and fail in interpreting the temporal information & blocks of text.

In order to deal with temporal or sequential data such as the sentences, we require algorithms that learn from the past data and also the future data in the sequence. Luckily, the Recurrent Neural Networks do just that.

Recurrent Neural Networks 

Recurrent Neural Networks are networks that are designed to interpret temporal or sequential information. RNNs use other data points in a sequence to make better predictions. They do this by taking in input and reusing the activations of previous nodes or later nodes in the sequence to influence the output.

Source

As a result of their internal memory, Recurrent Neural networks can remember vital details such as the input they received, which makes them be very precise in predicting what is coming next. Hence, they are the most preferred algorithm for sequential data like time series, speech, text, audio, video and many more. Recurrent Neural Networks can form a much deeper understanding of a sequence and its context compared to other algorithms.

How Do Recurrent Neural Networks Work?

The base for understanding the working on Recurrent Neural networks are the same as that for the Convolutional Neural networks, the simple feed-forward Neural Networks, also known as the Perceptron. Additionally, in Recurrent Neural networks, the Output from the previous step is fed as an input to the current step. In most Neural Networks, the output is usually independent of the inputs and vice versa, this is the basic difference between the RNN and other Neural Networks.

Source

Therefore, an RNN has two inputs: the present and the recent past. This is important because the sequence of data contains crucial information about what is coming next, which is why an RNN can do things other algorithms can’t. The main and most important feature of Recurrent Neural Networks is the Hidden state, which remembers some information about a sequence.

The Recurrent Neural Networks have a memory that stores all the information about what has been calculated. By using the same parameters for each input and performing the same task on all inputs or hidden layers, the complexity of the parameters is reduced.

Difference Between CNN and RNN 

Convolutional Neural Networks	Recurrent Neural Networks
In deep learning, a convolutional neural network (CNN, or ConvNet) is a class of deep neural networks, most commonly applied to analyzing visual imagery.	A recurrent neural network (RNN) is a class of artificial neural networks where connections between nodes form a directed graph along a temporal sequence.
It is suitable for spatial data like images.	RNN is used for temporal data, also called sequential data.
CNN is a type of feed-forward artificial neural network with variations of multilayer perceptron’s designed to use minimal amounts of preprocessing.	RNN, unlike feed-forward neural networks- can use their internal memory to process arbitrary sequences of inputs.
CNN is considered to be more powerful than RNN.	RNN includes less feature compatibility when compared to CNN.
This CNN takes inputs of fixed sizes and generates fixed size outputs.	RNN can handle arbitrary input/output lengths.
CNN’s are ideal for images and video processing.	RNNs are ideal for text and speech analysis.
Applications include Image Recognition, Image Classification, Medical Image Analysis, Face Detection and Computer Vision.	Applications include Text Translation, Natural Language Processing, Language Translation, Sentiment Analysis and Speech Analysis.

Conclusion 

Thus, in this article about the differences between the two most popular type of Neural Networks, Convolutional Neural Networks and Recurrent Neural Networks, we have learnt the basic structure of a Neural Network, along with the fundamentals of both CNN and RNN and finally summarized a brief comparison between the two of them with their applications in the real world.

If you’re interested to learn more about machine learning, check out IIIT-B & upGrad’s Executive PG Programme in Machine Learning & AI which is designed for working professionals and offers 450+ hours of rigorous training, 30+ case studies & assignments, IIIT-B Alumni status, 5+ practical hands-on capstone projects & job assistance with top firms.