machine learning and ar

Neural Networks 101

As VR and AR trends are experiencing rapid growth now, people want to know what it actually is and how it works.
The Question
As VR and AR trends are experiencing rapid growth now, people are questioning what it actually is and how it works. The first question is quite easy to answer, while the second one takes time. If you type AR or VR in a search engine, you will surely notice a term "neural network" next to them. Which brings up even more questions,doesn't it? We analysed queries about neural networks and came up with the list of most popular ones. Our Lead developer of Say Yes! app, Vadim, found 30 min for us to answer the questions and introduce you to the world of neural networks
What is a neural network?
Well, neural network is a layer-based automated system for giving answers to users' questions. For users it looks like inputting some data and getting a reply.

First, we train the system by giving it some information we know. For instance, we have 200 question and we know 200 answers for them. Questions and answers must relate to the same field.

We create a random neural network, input 150 of those questions, and get 150 answers from the system. After comparing the system's answers with the right ones, we learn that 10 out of 150 of them are right, all the rest is wrong. In accordance to some certain rules we change connections between neuron layers.

We repeat this action many times, until the number of right answers given by the system is satisfying. We can examine the network with those 50 questions we put aside at the beginning. In case the ratio of right answers given by the system is acceptable, the neural network is considered trained. For example, the simplest task is to determine if the figure is positive or negative. If we trained system to detect {1,2,5,10,50} as positive and {-1,-15,-25,-40} as negative, it will be able to answer which one is (-7).
Is it true that neural networks developers are trying to recreate a model of human brain?
At some point yes. The concept of neural network is taken from knowledge of brain neurons activity. We thought that if there is a natural way of data processing with no formal rules, why not using it as a basis.

Then it developed into something independent. No one is trying to recreate a brain, modern computers simply do not have such capacity. Theoretically, quantum computers will be able to simulate a human brain in future, but I haven't heard of any proofs of such computer working. Though Google and IBM have already announced they created and successfully tested a prototype of such machine, no one was allowed to try it.
Neural networks are now successfully used for image recognition (like in our apps) and image modification (MSQRD and Prisma)
Is mathematical neuron similar to biological one?
At its simplest - yes. There is a neural system, in which each neuron is connected to some others. Depending on input information, some neurons get activated and send the output result to us. If we need to categorize something ( e.g. to determine which quadrant of coordinate axis the point refers to by its coordinates), we have 2 inputs( X and Y). These are outcome neurons, which we send the information to. Also there are 4 outputs (coordinate quadrants). There are other neurons among those four. While processing the signal from input to output the neural network decides which of 4 output neurons it should activate. Human brain works in a similar way. Our eye gets some light of some certain brightness. Depending on this brightness, the eye needs to be opened, shut or narrowed. Central brain structure is much more complicated and there is no direct way signals go through. But formal neuron term is same, it is some certain unit in some certain state, connected with other similar neurons.
Will the 'games' that use neural networks be able to assist people in their everyday life?
There is a massive area of modern machine education, which includes neural networks and many other algorithms. Different algorithms are united into complicated systems, where it is hard to tell what is a neural network, and what is not. During last couple of years Yandex and Google fully transferred the search engine output task to neural networks. Google also added machine training to its systems work. It was in charge of electricity distribution, and saved 15% of systems consumption. Autopilot cars also have a neural network basis. Machine training is needed at the point where a human cannot detect laws and connections and describes the rules.
Augmented and Virtual Reality technologies are closely connected to neural networks and couldn't operate effectively without them. Even photo apps like Prisma are considered AR apps, as they change the way photos look.
Why has the neural network topic been so hot lately?
The idea itself appeared after the Second World War, and the first attempts to turn it into something real were made in 70s. Moreover, a new type of neural networks (the one that's currently used in Prisma app) was found out in 1998. By now we've finally learnt how to use the system and started to teach it, developed computers of needed power capacity, and found more ways to use the concept in practice. We also use the same technologies of image recognition in our Nailsmania, Shop 4 Rings and Say Yes apps to ensure smooth user experience while choosing a product to try on.
Did you like this article? More to come!
We're updating our blog weekly. Subscribe to get posts!
Made on