Artificial intelligence and machine learningDeep learning, neural networks, artificial intelligence and machine learning. These are the terms usually used for advanced analyses of large amount of data, which are flooding the world in the recent years. Technically, these terms stand for rather complex data models, that are able, for example, to recognize what is in a picture, to understand human speech or to diagnose a defective engine.
Advantages of deep learning
Compared to more simple statistical models, a deep neural network is able to recognize a person or other object, which is very useful in business or government applications. The value of recognition of a thief, of a defective part of an engine or of a picture may be immense; these technologies may help you to prevent losses or to multiply profits. A faster service to customers or to citizens, an immediate repair of a production line or detection of a perpetrator may save great amounts of money and make our lives more comfortable. It may be, therefore, useful to realize where the deep learning could help you. Do you want to simply count the products on the production line with a camera? Do you want to recognize the obstacles in the way of a robot? Welcome then in the 21st century, in the world where only the extremely clever ones can win.
What is a neural network?
Neural network is more complex than the common statistical models, because it has a greater number of layers. Although the first experiments were started already in the nineteen-sixties, they generally spread commercially only around 2010, together with easily available high-efficiency hardware.
With a little bit of a simplification, it may be said that a neural network is such software that is able to learn in a way similar way to human brain.
Let´s look at an example. We have pictures of elephants and of automobiles and we tell the model: these are automobiles and these are elephants. Individual layers of the neural network first recognize individual lines, then the trunk, and in the end they recognize the elephant as a whole. In every further photograph the model will be able to distinguish an elephant from an automobile. The same goes for sounds or other complex data.
The phase of such a model “learning” is rather demanding as to time and hardware in comparison with more simple models.
In a way similar to humans, some networks (“brains”) are more suitable for sound processing and other for processing structured data. For successful solution of concrete tasks, it is therefore useful to build specialized networks.
What is a deep neural network good for?
The models can learn almost anything. The only precondition is a sufficiently large amount of data from which the model should learn. At present, such an approach is most often used for analyses based on video or audio data.
The use of neural networks is customary in the following domains:
- Recognition of objects and people (eyes, face features)
- Diagnostics of appliances based on sound
- Understanding texts, processing of speech (NLP)
- Medical diagnostics (CAD)
- Production and maintenance of appliances (predictive maintenance)
- More complex predictive models in marketing
- Personalization (RTB and personalized e-mail messages)
- Data protection (antivirus programs and their databases)
- Web searching (Google Search Engine)
- Financial transactions (stock market forecasts)
- Detection of fraud (money laundering detection)
- Intelligent cars (Tesla)
- Government sector (car-plates recognition, recognition of persons, etc.)