Presentation at Ukrainian Blockchain Day 2017

Hello everyone! We’re glad to see you here, on Ukrainian Blockchain Days Today we’re going to tell you about how miners can earn on useful processing

Our company implements two technological frontiers at once: one of them is the blockchain, the other one is the deep neural networks We unite it in one platform Now we’re going to show you a very short explanatory video, and then we’ll tell you what we’re going to do, to make an enormous mass of processing power move from processing the abstract algorithms in blockchain to solving the fundamental problems of mankind Research in artificial intelligence started in the middle of the XX century Artificial neural networks are one of the oldest tools in the arsenal of artificial intelligence Their seventy-year long history has seen several boom and bust cycles, but ten years ago we all witnessed the deep learning revolution

Thanks to recent new ideas, neural networks have significantly improved state of the art results in speech recognition, machine translation, and many other tasks, including even superhuman performance in the game of Go New applications for deep learning emerge every day, and more and more talented researchers and creative entrepreneurs add to the world of artificial intelligence startups Modern deep networks contain a huge number of artificial neurons They have to be trained on huge datasets Labeling such a dataset is a lengthy process that requires prohibitive amounts of expensive human labour

The Neuromation company presents a solution that allows to generate and label data in a fully automated way with perfect accuracy and high speed, which solves the problem of lack of data Synthetic data are generated based on modern 3D graphics visualization algorithms, which lets us imitate any camera angle, any lighting conditions, and any surrounding environment with the necessary degree of photorealism Synthetic data have already been tested in practice and are successfully being used in our image recognition project for retail, where the problem is to detect and recognize different items on a supermarket shelf However, training modern deep neural networks suitable, for instance, for object recognition is a hard mathematical problem, very computationally intensive It may take weeks to train a modern large-scale neural network on a desktop PC with two modern GPUs, Where can you get this kind of computational power? A huge number of GPUs are currently being used to mine cryptocurrencies They earn much less money for their owners than modern AI-related companies spend on renting computational power from cloud-based providers

We are developing a Neuroplatform that implements blockchain-based smart contracts Our platform will allow to rationally use available computational power from the mining farms for cryptocurrency mining and redirect these computational resources to useful computation in the field of neural networks The additional functionality of the platform will unite designers, developers, researchers, and entrepreneurs, allowing them to share prepared data, including: 3D models of objects, labeled images, and trained models The platform will also provide services for setup and development of neural networks, labeling already existing data, and other works related to deep learning Neuromation – Knowledge Mining Platform And now I want our Chief research officer – Sergey Nikolenko – to tell you in more detail about in what way we use mining processing power, and why our company started to implement such unusual concept

Our report is divided into two parts: Sergey will explain the scientific part and then I will tell about our neurotoken economy Hello everyone, I’m glad to see all of you today I’m going to tell about our ideas of synthetic data while training neural networks, but since this conference is about blockchain and people here are focused on successful investments and how to earn more, I start with information, how this money will appear in our platform We’re living now in the age of a great machine learning revolution It’s already been 10 years since the deep learning revolution came, you must heard these words, but may not fully understand their meaning

Deep learning – is the deep neural networks training Neural networks is one of the oldest, classic and well-known instrument of machine learning And, to say the truth, many of architectures we use now, many of the basic ideas were already created in the 1980’s But in the 1980’s there wasn’t the processing power to train them, and most importantly, there weren’t data on which they could train And when in the 2000’s some new ideas of machine learning were invented, and, most importantly, Moore’s law drew up with the neural networks and allowed the really big models to be trained, this all got started

And now, as you heard it from the video, the most significant results in many fields were achieved by means of deep neural networks At first we worked only with the deep neural networks too, I’ll tell in details what exactly we worked with, it’s also a very interesting idea As Maxim said: “Neuromation is an unordinary startup”: we have two ideas at once And one day we faced the need of immense processing power And the common solution for any AI-related startup, that has just appeared, is to turn to any cloud-based service, for example Amazon Web Services, which is most typical and well-known service, and to rent there machines with GPUs, which are suitable for neural networks learning and for other computations

When we began to do it, the thought suddenly occurred in our mind, that Amazon consumes all our money, and their service was questionably expensive The rent of machine, equipped with 6-8 GPUs, is about $3-4 per hour on Amazon Web Services And then, Neuromation company came up with the second idea: we understood, that these GPUs are available in a quite large number in you, miners And miners earn $7-8 per day from the farm, that we have to rent for $3-4 per hour And there is a quite evident business opportunity here, and we’re going to use it in terms of Neuromation platform But now I’m going to tell you how we came to it, why we needed such immense computing power

I won’t go into the machine learning, the conference is not about it, but it’s important that there are two things needed to train the modern, vast models and deep neural networks, weights of which can amount to millions, and which can solve such significant tasks as image recognition One of these things is an immense computational power to train them, and the other one, the most important, is the lots of data, on which the deep neural networks need to be trained The more weight model has and the more complex it is, the more expressive it can be, and more things, interesting things, it can do, But on the other hand it means that it is more difficult to train, and more data needed to make it works So when we started to automate some industries, for example our first project is retail industry automation, that successfully continues up to this moment, We signed contracts with the large retail organizations, that need to solve very simple task, by today’s terms – to recognize different items on a supermarket shelf How that happens: the merchandiser comes and takes a photo of the shelf, and check item availability, comparing it with how they are supposed to be

Merchandiser checks whether everything is all right, or the items should be added there etc And during the stage of technology development, the neural networks have to handle this task, it’s merely a search for images on a photo, similar to Facebook that search for your faces on a photo, and this task seems to be more simple than that But it suddenly turned out that there are great number of objects on the shelves, and 170 000 of different objects must be recognized It’s quite a big number And, particularly, it means that lots of training data are needed there

The big selection is needed there, where each bottle or juice box is labeled and marked, and in order to train the network, it must be shown where there is a right object on a photo And it’s estimated that to train to recognize the 170 000 of objects, about 1 billion of labeled images are needed How can we get such number of labeled images? Up to this moment, the data labeling was basically a long, manual, and kind of painful process It starts with taking of photos, that in a case of dozens of millions of them is a sort of challenge But let’s imagine you have these photos, and then you come to some crowdsourcing service, such as Amazon Mechanical Turk, where you hire hundreds of thousands of unknown Indian and Chinese workers, who scrupulously label for you these photos for $2/ day

Unfortunately, even taking into account, that they work for $2/day, if we multiply it by millions, we still receive millions of dollars It means that such labeling type is merely inappropriate for really large tasks It has been done only one time in order to get the well-known dataset – Imagenet – which serves as a training base for all networks in Computer vision This project has been done only 1 time and it took several years, and had to be rechecked, because it’s possible that the Indian and Chinese workers can make errors It’s kind of not realistic to do such project again for each new industry

There was a need of more universal solution And what have we done: we realized that we can input the synthetically generated data to train the networks in Computer vision We can input 3D renders of all these bottles and juice-boxes It turned out that, first of all, it’s very successful idea for our retail project, because the items on the supermarket shelves are quite simple as 3D models, I mean that it doesn’t seem likely to do that with human face, comparatively with the Pepsi-Cola bottle Secondly, there’s still a task of so-called Transfer learning

We have to learn how to train a network on one data type, on some virtual images, and subsequently train them to recognize the real photos We managed to solve this task successfully It turned out that it is possible You can imagine that you have been shown a cartoon, and there was a new type of bottle, will you be able to recognize it on a shelf? Of course, no doubt The difference between cartoon and real photos is not the difference you use to subsequently recognize this object

So it’s obvious that it is possible It seemed to be difficult, but we’ve done it, and we learned how to train networks using the synthetic images And it was the single con we’ve faced, but there are lots of pros instead But the main pro is that we get free labeling with 100% accuracy The 3D render of a scene was made by ourselves, and of course we perfectly understand the location of all objects, we understand which pixel belongs to which object on that image, and the most important, we get labeling, that no Indian or Chinese is able to do

That means, we understand the object depth, their distance from a camera, at what angle they are on the shelf And all that is the additional information, which is impossible to get from manual labeling, and which naturally comes from the synthetic data And it goes with 100% accuracy As we created this by ourselves, we know everything And there’s no need to recheck the work of Indian/Chinese, there’s no need to make consensus protocols, when three workers need to confirm the same object, in order to put it into database

There’s required some manual work in creating 3D models, and after this you get an endless source of rationally generated synthetic data that can serve as a basis for training the neural networks And here we’re approaching to big requirements for the computational power, which we have encountered There’re two key components, which require this computational power: the first is generation of synthetic data, there’re needed GPUs for rendering, and it takes some considerable time for GPU to make a realistic image of high quality, it takes not the milliseconds, but rather up to the dozens of seconds So, to generate millions of such images we need either lots of GPUs or lots of weeks to work with them On the other hand, we need to train the neural network on the data we have generated, which is more complex computational process

That means that for the synthetic data and for the basic requirements for training of pattern recognition, as in retail industry, we needed lots of computation, and we needed lots of time, and great amount of GPUs And it was the period when GPUs were impossible to buy, it was about 4 months ago And at that time NVIDIA produced new GPUs, which we wanted to use, and I can’t say that they were expensive, but it was impossible to buy them, neither in Russia, nor in Ukraine, nowhere And in that period the idea of creating that platform came to us, the idea, that there are people, who already have these GPUs and who use them 10 times less effectively than we could use them Moreover, they use them to search for some hash collisions and we can employ these GPUs in more useful tasks for mankind And before I turn the floor over to Maxim, who will tell you about the platform and the project status, touching ICO and so on, I want to say that nevertheless I was telling you about retail, it was our first big contract project, and, actually, the same synthetic data can be usefully employed in many different fields We have already Retail Automation Lab, I told about it, and now we’re opening Pharma-biotech lab, tell you a little bit more about it, and Enterprise Automation Lab, it’s also about applying of synthetic data, which concerns about robot training, for example, robots, industrial robots, drones, self-driving cars, because nevertheless you can create a dataset for them, you can’t create it for any situation, you can’t launch million of drones, and make 500 000 of them crash in order to have enough data to train not to crash on them

I mean that you can’t simulate in reality some uncommon situations, and the synthetic data can help you in this Synthetic data allows you to simulate numerous different situations for drone, for example The particularly interesting field, where it can be used, is the medical industry Our second large project is going to start, the project is about neural networks creation, and their training for Smart cameras For example, it can be looking after babies

Let’s imagine the baby, lying in a crib, and the smart camera, watching it and defines whether everything is all right, won’t it broke its leg accidentally, whether it breathes, or whether it needs any help, and its current status at all And here we face the same situation: of course you can record the babies, but you can’t take millions of babies, and make 500 of them stop breathing, it’s not a good idea So again, synthetic data are needed here to simulate different uncommon situations, and we began to work on it And now I let Maxim Prasolov talk, who is going to tell you how it refers to blockchain, and what we do with that platform When we found out that useful processing is not only capable to solve the fundamental problems and be more profitable for the computational power owners, we realized, that by offering it to miners – which we did, and now there are about thousand of GPUs, making the useful processing for our platform – we realized that we can extend it to all processing power miners have

The previous speaker, Maxim, explained to you the things, which refer to mining inefficiency, emergence of new protocols, transferring to Proof-of-Stake, and basically all these indicate that the mining business will have to retain for something else, and we offer kind of diversification for mining business It doesn’t mean that tomorrow all farms will switch to useful processing The Blockchain auction, that we’re creating, will allow them to switch on/off the nodes, depending on task appearance If node is suitable for some task, such as generation of synthetic data, and it will engage in pool, that produces billions of high-quality images, or compute and train the networks with computational power, it takes the task and consequently receives a reward through smart-contract The reward is in neurotokens, which we now extend in the ethereum blockchain, but we understand that in the future we’ll need the faster blockchain, that does not exist yet

We expect that the transaction quantity growth on our platform will be up to 20 000/sec, because we’ll have to engage hundreds of thousands of different nodes to perform the useful tasks Such blockchain is only in the development stage, it doesn’t exist yet And there are several groups that are working on the development of such blockchain One such group is IOS, the other is Auterium, they try to create the fast blockchain, that will allow us in our system to implement a quite plain protocol of task distribution for miners

When we realized that there is no such blockchain, we chose a clear roadmap for us, to engage the big number of nodes and mining pools, which are currently negotiate with us and stand in line for engaging in useful processing, because we pay 20-25% more, than miners earn now with GPUs in mining light currency, such as Zcash, Monero etc We believe that this idea has not only existed, the market movement confirmed it You must heard that Bitmine produces Smart farms, that means, universal rigs, that can handle with cryptocurrency producing and with neural networks computation at one time The Smart farms is the Future And we’re creating the platform, that will distribute the tasks, and help the Smart farms owners to earn

We called it Knowledge mining And this platform is in the development stage, we are extremely close to the computation distribution protocol In February we’ll show the first prototype of the platform Currently for some our clients we’re doing it manually In the future, the useful computation will be the most important economic challenge for processing power owners

We often discuss that without cryptocurrency, it has to be created merely in order to provide the immense computational power It was a severe bottleneck for AI development in the world But it was solved because the immense computational power is in the private hands, and our task now is to create the single protocol for it, to produce really useful things We even have a story that AI named itself as Satoshi, and created the Bitcoin to make people invest in the computational power It may sound fanciful, but we’re on the technology frontier

The technology pattern is changing There was a time when Andrew Ng, the cofounder of Coursera the cofounder of Coursera and the important person in Deep Learning industry, said: “AI is new electricity” The Intel director said: “The data is new oil” We live in the world, where there are new currency, new oil, new electricity, and breakthroughs are possible here Our company figured out a way to synthesize the oil, e

g to produce the data in unlimited quantities Our approach is gaining strength around the world Our token sale will begin less than in a month, on November the 28th Now we’re having pre-sale

We’re very careful when choosing the investors Not everyone is registered, because each of them goes through the KYC and AML procedures We know who our client is, and we work only with the reputable investors We don’t need a large capital to launch the platform We understand exactly how much money we need to make the platform work You must know this term “cosmonauts”, and we’re not the “cosmonauts”, we’re down-to-earth people, who work with clear and real technology

I mean that neural networks and blockchain are the technology more of real sector than of virtual sector Industry automation will include any industry, from retail to resource extraction A lot of professions will disappear, but a lot of new will appear instead Our platform creates a certain set of business processes, which can generate liquidity for users of this platform as well as for those who provide resource for this platform It includes synthetic data generation, labeling data, producing the architectural models, which train on the data, and the model training itself

The components of the system will be united in a single clear user interface And although our scientists say, that it will take a long time before any housewife will use our platform to train a neural network for her garden or house, for example, but we understand that we’re moving toward liberalization of access to this technology In the near future you will be able to create your own neural network, which you will need, in the same way, as now you use WIX service to create your own website There is no doubt that you will need the neural network, as the smart assistants will be everywhere It is important that our platform has one key thing in its development: we believe that economic subjects will emerge without human, there will be a lot of non-human economic subjects, such as a bank account of your self-driving car, of your smart-house, of any device, even a bank account of your node, which will sell its power on our platform

It is known as machine-to-machine economy And we’re creating a foundation for it Eventually, drone, robot, self-driving car or any other device will be able to sell knowledge to the common system, to the neural networks, as well as receive this knowledge from neural networks By 2025, as the researchers of European Commission predict, there will be about 25 billions of devices, such as drones, robots, self-driving cars And these devices will create a new type of economy

And when we say about the future of cryptocurrency, it seems obvious, that it will be an equivalent of service exchange for machine-to-machine economy I think few people heard this, but in several years everybody will talk about it, similar to how people talk about blockchain today We’re creating our platform on the basis of blockchain, because it is a single technology, which is able to handle with quite difficult and great number of transaction in system Let’s imagine that our nodes, which we rent, can be located in Chine, in New Zeeland, in the USA, wherever We can’t sign a contract with each of them at once, especially if we need them only for 1 hour or for couple of minutes

We need that computational power to be turned on, then it solves a certain task, then receives a reward, and then it should be turned off to switch to other tasks, such as cryptocurrency producing, or any tasks, which are related to everyday life Blockchain for us is completely proper eco-system for creation the smart-contracts system, for paying nodes for useful computation, and for combining this in one system Out team consists of people, each of whom has succeeded in business, has built a career We have a wonderful advisor team, one of its members is Andrew Rabinovich, who holds a key position in Deep Learning He developed and created the algorithm Google Image, which allowed you to search for photos

It's also David Orban, a cofounder of the Singularity University, who joined our team, because he was inspired by our idea Knowledge mining is our ambassador throughout the world Constantine Goltsev, an investor and a cofounder, and my partner, who created numerous startups all over the world Denis Popov, who was a CTO in Viewdle company, and who improved Google algorithm for face recognition Fedor Savchenko, is our CTO, who creates the cartoons we show to AI to train it

Our team is composed of strong, serious and intelligent people, who understand that realization of such market is our huge future And we’re ready to share it with you We’ll be here all day, and feel free to ask us any questions, you can come to us if you are miner, if you need the neural network algorithms, or you’re just interested to engage in Neuromation economy Thank you for your attention

Comments are closed.