The brain is one of the most complex and vital organs of the human being. It has been estimated that the human brain contains 50 to 100 billion neurons, of which about 10 billion are cortical pyramidal cells, which transmit signals to the body through a billion trillion synaptic connections.
Machines are beginning to use the ‘brain’, much like that of the human being. An artificial neuron network is a mathematical tool that models, in a simplified way, the functioning of neurons in the brain. The architecture is as follows: the neurons are organized in layers, where the green neurons are the inputs that receive the information, the blue ones are the hidden ones, which contain intermediate calculations of the network, and the yellow ones are the outputs that contain the result. There is usually one input layer, one output layer, and several hidden. The more hidden layers, the more complex the network and the better the predicted results, but it is also more difficult to create the model.
All neurons contain a number called bias that indicates their importance in the network, in addition they are all connected to each other and contain a weight that indicates the importance of the connection. The operation in the network of neurons would be simple if it were not for the activation function , which receives as input the sum of all the numbers that arrive through the previous connections, transforms the value by means of a formula , and produces a new number. The mission of this activation function is to keep the numbers produced by each neuron within a reasonable range.
There are different architectures of different neuron networks , among the main ones we have: the restricted Boltzmann machine (RBM), the deep belief network (DBN), the convolutional neuron networks (CNN), and the recurrent neuron networks (RNN) .
Deep learning: applications
The uses of these algorithms can be divided into four groups: word processing, image recognition, object recognition, and speech recognition. These algorithms are of great importance in the design of the software in autonomous cars, face detection in Facebook photos, or in the new Google translation algorithm.
If you are interested in creating your own models of neural networks, an alternative is to use a platform such as H20.ai or Dato. Various text processing or object recognition algorithms can be applied in them relatively easily. The other alternative is to use libraries like Theano, Deeplearning4j, Torch or Caffe. Although my preferred alternative for working with neural networks is the Google library, TensorFlow, which can be implemented in Python.
According to the Gartner cycle for emerging technologies, deep machine learning is and will be one of the areas that will need the most demand from professionals in the coming years.
Robert Brunet is a researcher in chemical engineering at the Rovira i Virgili University (Tarragona)
