You are on page 1of 2

Deep Learning

Deep learning is a subset of machine learning, and machine learning is a subset of artificial intelligence.
To understand the concept of Deep learning, one must understand its building blocks - Artificial Neural
Networks (ANN). In simple words, ANN is a machine learning technique/model for analysing data
through a network of decision layers. The naming of this technique is inspired by the algorithm’s
structural resemblance to the human brain neuron.

An ANN consists of interconnected decision functions, known as nodes, which interact with each other
through edges. The nodes are separated into layers namely input, hidden and output. Data (image,
text, audio, etc.) is first received by the input layer, where features are detected. The hidden layer(s)
then analyse and process the input features, and the result (prediction) is shown as the output layer
(refer figure).

Larger the number of hidden layers in an ANN, better is the model’s capacity to analyse complex
patterns. Hence, when patterns in the data become overly complicated or the data has high number
of input features (for e.g., image pixels), a model with a deep number of layers is needed for superior
processing abilities. This is known as deep learning. What makes deep learning “deep” is the stacking
of at least 8-10 node layers.

Object recognition, as used by autonomous cars to recognize objects such as pedestrians and other
vehicles, is a popular application of deep learning and it uses upwards of 150 node layers. Other in-
demand applications of deep learning include image classification, speech recognition, text processing
& synthesis (also known as Natural language processing or NLP) and time series analysis.

Based on the above-mentioned popular applications, the two most widely used deep learning models
are:

 Convolutional Neural Networks (CNN)


 Recurrent Neural Networks (RNN)

The use cases and the commonly paired deep learning technique are tabulated below:

Popular applications CNN RNN


Text Processing 
Image Recognition 
Object Recognition 
Speech Recognition 
Time Series Analysis 
Classification 
Let us now compare the major features of CNN and RNN. From the figure below one can easily
understand that both CNN and RNN are modified versions of ANN.

The CNN has unique layer structures in the form of convolution (filtering) and pooling layers which are
dedicated to the task of feature extraction. This helps in finding the key features in data and ignoring
the trivial ones, thus getting more accurate predictions. This property of CNN is essential for image
and object recognition.

The difference between RNN and CNN is that RNN not only passes data forward throughout the layers
but also feeds the data back into itself. In case of CNN the data passes only one way forward. This
means RNN can remember the context before and after the current word in a sentence for example.
Thus, RNN is well-suited for sequential data, such as text and audio.

I hope this article could help the readers in improving their general understanding of deep learning.

You might also like