What is deep learning?

Question

Reading one of the answers, about Algorithm to detect nudity with good accuracy, I have seen that a term that is much quoted and referred to is Deep Learning. For me, translating Deep Learning didn’t make much sense.

• What is Deep Learning?
• Where, what or who studies Deep Learning?
• In what Deep Learning influence?

Answer 1

Deep learning is associated with artificial intelligence and mainly allied to neural networks, the goal of a neural network is to learn and recognize patterns, for this it makes use of several neurons, which form a network, as the following image.

The W values represent the input of data in a neuron and the output of it will be linked in other neurons, each neuron has a "weight" which is called threshold and which defines the function made by it.

Important to point out that they are unsupervised networks, they learn alone without human supervision.

Deep learning-inspired machines can learn complex abstractions of data through a hierarchical learning process very similar to what happens in the human brain - especially in the visual cortex.

Soon a network is assembled and the neurons in the inner layers are responsible for identifying small patterns in something larger, such as checking character "7", a part of the network recognizes the angles and directions of each pixel in the image.

In his cited case, of pornography, the neural network is able to identify small patterns and that if summed can affirm that, for example, what appears in the image is a penis. There are several techniques of evaluation of the neural network, I just believe that this is a little outside the scope of your question, however I suggest reading on Perceptron and Backpropagation.

Deep Learning can input various things, mainly unstructured data and can be used for photo, sound, video, text and other recognition.

For more details about Deep Learning I suggest that one posting.

Answer 2

Deep Learning are artificial neural networks with many layers. These layers when cleverly combined resulted in great advances in the area of artificial intelligence.

Usually in machine learning problems it is necessary for a human being to determine Features that describe a data set. A Feature is, for example, what Luiz describes about color histograms in that reply. The great advantage of Deep Learning is the ability of these models to extract Features raw data, without the need for human intervention.

Although simple models like a 3-layer Multilayer Perceptron are already considered as Deep Learning, is not there its great success. What has really revolutionized this field are more advanced network architectures like:

• Convolutional neural networks: for image recognition. Example: Street name identification in Google Street View
• Recurrent neural networks (such as LSTM’s and GRU’s): for sequences in general (text and sounds). Example: Google Translator
• Convolutional and recurrent network combinations: for video processing. Example: Thumbnails of Youtube videos
• Neural Networks with attention that seem to be less specific. Read here

And many others that come up every month. From the examples I showed, you can see how much Google uses deep learning in their research projects, but it’s not the only one. Just log into the site of the research group of other major technology companies:

• Facebook
• Baidu

And many others.

Deep Learning is very influencing the area of computer vision. Don’t have a self-driving car company, for example, that doesn’t use deep-Learning as the central technology of automated driving software.

Currently, the best framework to start dealing with Deep Learning is the keras. So much for python how much to R.

Remember that Deep Learning can be used in all forms of machine learning: unsupervised, supervised and enhanced learning.