Introduction to Cindicator (CND) Part 4

Deep learning neural networks.


Deep learning neural networks are different in depth from other types of neural networks, meaning that data on deep learning networks passes through a bigger number of node layers.

Early versions of neural networks would typically have just one layer of inputs and outputs. Today, most experts and organizations would define a machine learning network as a deep learning network if it has at least three layers, including inputs and outputs.

On a deep learning network, each layer of nodes becomes an expert on a certain set of features that are based on the results of work of the previous layer. The deeper into the layers of a deep learning neural network, the more complex the features and the harder the tasks that nodes can perform. For example, with facial recognition layer one could be training on separating human faces in a stream of video or pixels of audio from all kinds of other objects and then learning about whether it identified a human face correctly. Layer two could be identifying certain parts of human face, such as noses, eyes, eyebrows and so on. Layer three could be matching faces against identity databases and establishing the identities of people.

Such a structure of layers is also known as a hierarchy of features. The hierarchy is what makes deep learning networks capable of handling very large amounts of data with billions of parameters that they may need to analyze.

What is critically important here is that a deep learning neural network could work with unstructured data, such as a live video stream. This is critically important because most of the existing data in the world and data that is incoming is not structured.

For example, when an autonomous vehicle is driving down the road, the stream of data that it gets via its sensors has no structure to it whatsoever. It is the job of a deep learning network to distinguish between traffic signs and traffic lights, humans, other moving objects, and non-moving objects in that stream of data. Also, while is it obviously possible for a human to analyze all this data when a human is driving a car, it is not possible for a person to analyze the feeds from several cars in real time and make decisions about what the cars should do, when they should stop and when they should go, accelerate or slow down, and so on. For these reasons, as the amount of data in the world is increasing, neural networks and deep learning with play an increasing role in the existence of the world.

Here are some examples. A deep learning network could separate photos based on the presence of animal and people in them. This is what is already happening with so-called smart albums. The same principle applies to all kinds of other data. For instance, a deep learning network could separate spam emails from news alerts from complains of angry customers. The complains could go to specially trained representatives that know how to make customers happy again and how to solve problems. Emails with testimonials from happy customers could go into a different cluster.

If the data has the parameter of time, deep learning networks could separate dangerous behavior from regular behavior and healthy behavior from behavior that is exhibiting mental symptoms.

Typically, engineers would first train a neural network on labeled data similarly how parents or teachers would explain a child the difference between a dog and a cat. Each node on a neural network learns automatically by trying to analyze the sample inputs and increasing the probability of the guesses. This is similar to how a person that does not know the difference between dogs and cats may be occasionally referring to pictures of animals that he knows are dogs and cats when looking at pictures of new animals.

In the process, the neural networks learn to recognize correlations between certain data points and desired results. Next, engineers would feed unlabeled data to a trained neural network and give the network access to more data than they would in the regular machine learning process. This is what allows deep learning neural networks to perform so well and this is what is the key to the success: the more data a network gets and the more accurate are analysis and predictions of the existing data, the better will the analysis and predictions of even larger sets of unlabeled data.

Obviously, nodes on a neural network could be decentralized and the information on such a network could be stored on blockchain, making it immutable and tamper-proof.