Many know me as the author of the graphic prose, producer in multimedia, senior Manager, PR man, and it’s true, I was doing that. Worked with international retail and many industrial brands, commercial banks, investment and insurance companies.
I have a hobby â€” cartoons. This year my team did an adaptation of “Kobzar”. We got zombie horror cartoon. Since animation is a time â€” consuming process, in the course of work I came up with the idea: need to automate the process, to write algorithms and learning neural network using 3D animation. A bit of history
The neural network is machine algorithms in which one tries to teach a machine to reproduce the human nervous system. Back in the 1940-50s, scientists began to be interested in them. Built perceptrons in practice, in hardware and with their help tried to recreate the human brain, mathematical model of neurons. Then came the “Ice age”: the 1970s all forgot about them and switched to fiction in literature and film. In the 1980s, remembered again. And by the end of the 80s started to experiment and produce algorithms. To train the neural network did not work because not enough computer power, data and ideas.
The neural network in practice
The neural network can be described as a classifier and the method of segmentation of information, which not only acts in accordance with a specified algorithm and formulas, but also on the basis of past experience.
A sort of child that each time he folds the puzzle, making fewer mistakes.
To teach a neural network something to recognize, you need to provide an average of 5 thousand images of one object or subject. Each object in the image should be laid out and described: this man is the cat, it is a cube, and it is a dog in a sweater. Only then the neural network will learn to distinguish the umbrella from the dog, the car from a person, a bottle from a motorcycle. There is a paradox in machine learning â€” people learn the neural network by hand, and people make mistakes. This process is very expensive and time consuming. The neural network trained on synthetic data application of models on real photos. Someone was trying to teach Autonomous cars inside the computer game GTA. But we have simplified this process. What we came up with
While engaged in the cartoon, we figured out how to automate these processes. We created a 3D simulation of reality, and thanks to these models, learning of the neural network. Thus our design allows to achieve accuracy in learning networks, minimizing the human factor.
The network is trained on synthetic data, transferring them for real. And this process is known as transfer learning (transfer learning). Just 2 months ago, we have managed to achieve high quality recognition. How it works
We found people who were helpful in our development â€” that is, the automation of retail. Our first client was a vast network of retail and retail. For example, on the shelves in the supermarket are a huge amount of goods. But in order to help suppliers to control retail, we transferred it into the figure, by teaching the camera to see what is on these shelves. We “put on shelf” 3D models in any combination â€” is a 1 billion photo-realistic images.
If the employees did it all manually, it took 120 million man-hours.
In addition, retail automation, our design will help to test drugs because the neural network can simulate a living cell. Or you can do training industrial robots and drones in virtual environments. But first, for the implementation of these processes, we wondered, where to get resources for this. What we did
For the existence of Neuromation needed graphics card. Many graphics cards. At first, all opportunities for the training we rented for AmazonCloud and paid tens of thousands of dollars, for the use of their facilities. Then we decided to buy their equipment, but faced with an amazing phenomenon. It turned out that the graphics card we need processors-GPUs for neural networks, need for mining, for mining cryptocurrency. In the month of may, the speculators sold the card in troegerae. We searched for them in all markets, but we got a weak video card from the United States. For the miners couldn’t keep up â€” they filtered all the power. And we decided to offer them a deal. We have calculated how much they earn and found that one farm with six cards gets $7-8 a day (it was about the money which we pay with Amazon”). We went to the miners and asked them to make $10 a day in cryptocurrency. Just a day, we were asked to send addresses where to carry a card. Now we cooperate with a lot of miners who work for the good of mankind. What happened in the end
We have created a bridge between the miners and scientists, inviting them to earn more for utility computing. After a month we had a thousand cards that give enough capacity for our experiments. And two, we along with a team of scientists confirmed the hypothesis on synthetic data.
Now they say that data is the new oil, and we managed to find a synthetic oil.
Useful computations are much more profitable than to compute the abstract algorithms in the blockchain. Industry deep and machine learning are ready to pay miners more than they earn mining cryptocurrency.
Our research shows that all this can be done in the future with two things â€” our synthetic approach to date and the incredible power that was in the hands of people suffering from cryptocurrency fever.
So we created a platform for KNOWLEDGE MINING and announced a startup competition. The prizes will give technological power, which can train your neural network.
We have created a product with user-friendly interface, where any person, not versed in the nuances of machine learning, could teach the neural network to recognize all that can be recognized under its task.
Even if you have no desire to change the world, you just need to do what you like.
In my case, hobby cartoons that showed how it is possible to train the network with the help of 3D animation.