Deep learning is a sub-category of machine learning and is closely linked to artificial intelligence. It is via this process that machines learn to reason in a logical and consequential way, by using neural networks.
The development of increasingly complex algorithms has allowed for the ‘teaching’ of deeper concepts to the program, which, by ‘learning’ more information, is able to provide more precise, reliable forecasts.
Deep learning in detail
Neural networks are what distinguish deep learning from other branches of AI. It involves algorithms created on different levels and arranged in a cascade pattern. Each one has the task of carrying out transformations and extractions, which are then used by the next node and so on.
To further illustrate this concept, we might consider the phases in the simulation of a visual process. The first algorithm has the task of recognising the sides of a geometrical shape, the second interprets its basic shape (triangle, rectangle etc.), the third uses these rough shapes to construct more complex images and so on until a faithful reproduction of the real image observed by the machine is created.
This model of successive algorithms (via the addition of nodes to the neural network) allows the creation of very precise forecasts and simulations. You can picture it as an organic brain, whose calculation potential is proportional to the dimensions and number of neurones it contains.
Compared to machine learning, deep learning has no limits in terms of improvement potential. This aspect, in fact, depends entirely on the quantity of data inputted during training, designed to teach the machine how to make the results it provides more precise.
Supervised and non-supervised algorithms
The purpose of deep learning is to create thinking machines, which are able to reproduce reliable, logical thought processes. To do this, the intervention of a data scientist is required, a professional who is able to analyse and asses the data, while recognising the essential data needed to help the AI with its learning.
Cases such as this are known as supervised algorithms, in which human intervention is fundamental and functional to the machine in order for it to improve. There is also another type of algorithm, known as un-supervised, in which the artificial intelligence itself learns autonomously, based on previous experiences.
The learning process, in these cases, is more complex and is based on patterns. In practice the machine is capable of recognising similarities, relationships and cycles in the data it is analysing, so as to build realistic forecasts.
The data scientist is important in this case too, since he has the task of interpreting the results provided, while ‘explaining’ to the machine any evaluation errors committed, so that they can be avoided in future.
How deep learning can be used
The development of increasingly complex neural networks, supported by technologically advanced hardware, is the key to creating deep learning systems which can be useful in many different sectors. Below are some examples of its most common applications:
- Computer vision is the system found in smart cars (but also in drones and robots), enabling them to drive in total autonomy. Deep learning in this case is able to choose the best route by recognising shapes, dimensions and the distance between the vehicle and obstacles;
- Chatbots are another good example of how this technology can be used. E-commerce customers often do not even realise that their requests for assistance via chat are being answered by an intelligent computer;
- Facial recognition systems for video surveillance are based on a kind of neural network, capable of detecting facial features and comparing them with those stored in a specific database;
- In the medical sector deep learning is helping researchers make new discoveries about DNA and carry out ever more precise x-rays;
- For businesses it is an excellent system for gathering and analysing data, as well as being an indispensable forecasting method, which can be extremely advantageous.
Translated by Joanne Beckwith