Deep learning and neural networks

Today we focus on Deep Learning and an important algorithm part of the Deep Learning; Neural Networks. Deep Learning correponds to a category among the machine learning methods. These methods are characterized by applying several complex functions to some inputs in order to obtain outputs that solve a problem. If Deep learning is highly complex and difficult to understand, the interesting thing about it is that it is inspired from neuroscience. Our brain works the same way applying complex functions to electric signals coming from our nerves. One of the algorithms trying to copy our brain is the neural network.

Modelling a simple neural network
with one hidden layer

The first layer corresponds to the inputs modeling the signal we get from our nerves. The last layer corresponds to the outputs characterizing the solution of the problem the neural network is solving. The layers between are called hidden layers. Each cell of these layers takes the inputs of the previous layer as argument, applies a function to them and passes the value of this function to the next layer until the output layer. These cells represent the neurons of our brain communicating through synapses. What is remarkable about these algorithms is that it works pretty well with automatic speech recognition, natural language processing or facial recognition. These things are hard to solve mathematically, using equations but when we know a language, we easily understand someone talking to us, no matter how his pitch is or his pronunciation. It shows how closer we are getting to modeling a human brain and to reach Artificial Intelligence. 

For example, Neural Networks have been used since the 80's to decipher the amount written on the checks you give to the bank. More recently, thanks to the increased power of our computers, more complex neural networks can be trained with up to one billion neurons and tenths or hidden layers. Recently in 2014, Facebook unveiled an algorithm called DeepFace that can recognize specific human faces in images around 97% of the time, even when those faces are partly hidden. A human would have a hard time being as good as this algorithm.

For more illustrations of what can be done with these deep learning algorithms, I recommend you to watch an incredible video from TED where you can see Jeremy Howard, the previous President of Kaggle, a community and competition platform of over 200,000 data scientists. The last point he exposes is an unsupervised algorithms. In other words, algorithms which can learn without the need of human help. Algorithms were developed that way because helping the machine to learn by giving examples can be very costly in human work. These kind of algorithms are amazing because after going through lots of images, it can recognize photos of cats sleeping from cats jumping. Of course, without human intervention, the only thing left to do is to give the corresponding labels to the classified photos so that the computer can communicate with us with words.

Now these machines have no human behavior, but add this technology to glasses and we get reality-augmented glasses capable to give you indications about who or what is standing in front of you.