What is Deep Learning?

You are passionate about Artificial Intelligence and you want to better understand what Deep Learning is? This article is for you! 

Are you passionate about Artificial Intelligence? Are you intrigued by the fields of Machine Learning and Deep Learning? 

Here we explain what Deep Learning is.  

What is Deep Learning? 

Deep Learning is a branch of Artificial Intelligence (AI). Derived from Machine Learning, this technique consists of creating algorithms capable of learning autonomously. 

Deep Learning is based on artificial neural networks. 

Their development has led to significant advances in the analysis of audio and visual signals by computers, facial and voice recognition, and language understanding.  

In this article, we explain how Deep Learning works. 

Deep learning and the brain  

To understand Deep Learning, it is necessary to recall some neurological notions. 

The brain is composed of more than 100,000 billion neurons. Each neuron consists of dendrites, through which it receives electrical signals. It has an axon, through which the signal passes to access the synaptic connections whose role is to transmit the signal to another neuron via its own dendrites. 

These innumerable branches, making up a complex network with several layers, make the brain capable of collecting, analysing and expressing an almost infinite amount of data. 

Deep learning is therefore based on the development of artificial neural networks. The greater the number of neurons, the better the learning performance. 

How does Deep Learning work? 

Artificial neural networks are divided into three parts or layers: 

  • The family of input neurons through which the values enter. 
  • The family of intermediate neurons, the hidden layers, in which the values are modified. 
  • The family of output neurons from which the analysed data leaves the machine. 

To enable the algorithms to learn autonomously, engineers must first train them. During this first phase, called the supervised learning phase, we supervise the correspondence between inputs and outputs. 

Feed forward is a technique where we pass data to our input neurons. Each value travels from one layer of neurons to another in the network. During this passage, a connection weight is assigned to each value. It modifies the values in the network and recombines them. 

The activation function determines whether data will be transmitted from one neuronal layer to another. Thus, each neuron can be either active or inactive. Among the best known functions are Sigmoid, Relu, and Tanh. 

At the output, the values are therefore modified by the algorithms and may or may not correspond to the input values. In case of error, error functions allow the neurons to change the connection weights and thus recalculate the data throughout the network. These revisions allow the algorithms to output new values more or less close to those expected. 

Gradually, the algorithms become more refined and eventually are able to improve their in an autonomous manner. The more data and input values the machine analyses, the more accurate its automatic learning will be. 

To illustrate this, let’s imagine that we want to teach our algorithm to recognise an image of a dog. Over time, the machine should be able to recognise all images of dogs. In other words, when we enter data (pixels, shapes, etc.) corresponding to a dog image into the machine, the computer must produce the adequate value.  

Through the training phase, the algorithm will accumulate data related to the shape of a dog. After modifying, calculating, adjusting and rectifying the output a large number of times, the machine becomes capable of learning autonomously and perfecting itself. 

It is important to recall in this context convolutional neural networks. This is a technique inspired by the human visual system that allows a computer to recognise very subtle details. 

Deep Learning applications

There are many applications of Deep Learning. We can list some of them: 

  • Robotics 
  • Bioinformatics 
  • Artificial Intelligence 
  • Health 
  • Security 
  • Recognition of shapes, sounds, languages and text  

In fact, we are now surrounded by Deep Learning. Automatic text translation on Google, facial recognition on Facebook, voice recognition on YouTube, etc. 

Artificial Intelligence, Machine Learning and Deep Learning are real revolutions that are already changing our world. 

To find out more about Data ScienceTech Institute’s Applied MSc programmes courses, contact us here