Blockchain circle

One stop hot information platform

About us:

Blockchain circle provides the latest information about blockchain, digital currency, digital wallet, exchange, metauniverse, bitcoin, Ethereum, contract, financial management and so on, and always pays attention to the latest market...

Machine learning market based on blockchain

Time : 10/10/2021 Author : mc4kpt Click : + -
        To truly influence the world, blockchain has a very important breakthrough, which is the combination with artificial intelligence. For start-ups, the biggest bottleneck in the development of artificial intelligence is the lack of enough high-quality data. These data are almost owned by technology giants. Through the incentive mechanism, the blockchain forms an open multi-party win-win network of data providers, data scientists, users and node service providers. Once a positive cycle is formed, it has the opportunity to build the most powerful artificial intelligence system in history. Of course, there will also be difficulties in the implementation, such as the implementation of multi-party computing, privacy protection, value measurement of different contributors and other details.
        At present, there are many problems to be solved, and it seems difficult. However, as more and more innovators enter, the problems will be solved one by one. They combine two powerful elements: one is machine learning for privacy. It allows model training without disclosing sensitive privacy data. The second is the incentive based on the blockchain. The blockchain system will attract the best data and models to make it more intelligent. The result is that in the open market, anyone can sell their data while protecting their privacy. The developer obtains the best data required by the algorithm by providing incentives. Building such a system is challenging, but a simple initial version seems possible.
        I believe that such a market will evolve us from the current Web2.0 era of data monopoly by large companies to the Web3.0 era of open competition between data and algorithms. Both are directly commercialized. This idea came from talking with Richard of numerai in 2015. Numerai is a hedge fund that sends encrypted market data to data scientists to build their own stock market models. Numerai combines the best models and submits them to the "meta model", and trades in the market through the "meta model". If the model performs well, data scientists will benefit.
        It seems like a great idea to let data scientists compete. It makes me think: can we create a completely decentralized system that can be applied to more general scenarios? My answer is yes. For example, let's try to create a fully distributed system for cryptocurrency transactions. The following are elements of a system like architecture:. The modeler creates a model and selects data for training. Model training is carried out without disclosing the basic data. The model will also have an interest. After a period of time, the transaction generates profit or loss. This profit or loss is distributed among the contributors of the meta model, depending on how intelligent the model is.
        If the model causes loss, part or all of the interest in the model will be confiscated. At the same time, the data provider of the model will also implement a similar benefit distribution or equity reduction mechanism. The calculation of each step is either centralized, which is verifiable and challenging (using verifiable games similar to truebit), or decentralized, using secure multi-party computing. Incentives to attract data are the most effective part of the system, because data is often the most important limiting factor for most machine learning. Through open-ended incentives, bitcoin has created the most powerful emerging system in the world. Similarly, a well-designed data incentive structure will bring the best data in the world to applications.
        And it is almost impossible to shut down this system with thousands or millions of data sources. There is no open competition between algorithms or models before. Imagine a distributed Facebook with thousands of competing "information flow" algorithms. Data and model providers can see that they get a fair return on value because all calculations are verifiable, which makes them more motivated to participate. The multilateral network effect from users, data providers and data scientists makes the system rapidly self reinforcing. The better it performs, the more funds it attracts, which means more potential value output. This will attract more data providers and data scientists, who will make the system more intelligent, thus attracting more funds and realizing self circulation.
        2) Prevent the economic value of data and models from leaking. If it is not encrypted in public, the data and models will be copied and used for free, and these people may not have contributed any work. That is, there is a problem of free riding. Part of the solution to the free riding problem is private transaction data. Even if the buyer chooses to sell or release the data again, its value will decline over time. However, this method restricts us to the scenarios used in the short term, and still generates typical privacy problems. Therefore, a more complex and powerful method is to use a secure computing method. The secure computing method allows the model to train the data without leaking the data itself.
        There are three main forms of secure computing used and studied today: homomorphic encryption (he), secure multiparty computing (MPC) and zero knowledge proof (zkp). At present, multi-party computing is the most used in machine learning, because homomorphic encryption is often too slow, and the effect of zero knowledge proof applied to machine learning is not obvious. Secure computing methods are at the forefront of computer science research. Although they are several orders of magnitude slower than conventional calculations and represent the main bottleneck of the system, they have been improving in recent years. To illustrate the potential of private machine learning, imagine an application called the ultimate recommendation system. It will monitor all the operations you perform on the device: browsing history, all the operations in the application, pictures on the mobile phone, location data, consumption history, wearable sensor devices, SMS, home camera, camera of future ar glasses.
        Then it will give you advice: the next website you should visit, the articles you want to read, the songs you want to listen to or the products you want to buy. This recommendation system is very effective. It is more powerful than Google, Facebook or any other existing data island. Because it has the most complete view and can learn from the data more timely, otherwise the data will be too private to be used. Similar to the previous cryptocurrency trading system, it allows models that focus on different fields (e.g., websites Recommend Music) to compete to obtain users' encrypted data and recommend content to users, and even promote users to contribute data or attention through payment.
        Although Google's joint learning and Apple's differential privacy have taken a step forward in the direction of private machine learning, they still need to be trusted by users, and users are not allowed to directly check their security. There are problems such as data isolation. A simple construction of algorithmia research is to give a reward to the model above a certain backtest threshold:. Numerai has started further improvement: it uses encrypted data (though not completely homomorphic). It combines crowdsourcing models into meta models and provides rewards according to future performance (one week's stock trading time), instead of conducting backtesting through Ethereum token numeraire.
        Data scientists must use numeraire as a token in the game to stimulate future performance, not what has happened. However, at present, the most important part of its data distribution is limited. Others start by creating secure computing networks. For example, openmine is creating a multi-party secure computing network, which trains machine learning models based on unity and can run on any device, including game consoles (similar to foldingathome), and then extends to secure MPC. Enigma has a similar strategy. A fascinating end state is co ownership of the meta model.
        Meta models can give data providers and model creators rights, which is directly proportional to the intelligence of the model. The model will be tokenized, dividends can be paid over time, and may even be managed by those who train them. A common intelligence. The original openmined video shows the closest idea I have seen so far. My idea of evaluating the blockchain project is: in a certain range, from physical native content to digital native content, and then to blockchain native content, the more the blockchain native content, the better. The less the native content of the blockchain, the more it relies on the introduction of trusted third parties, increasing the complexity and reducing the ease of use built with other systems.
        This means that if the value created is quantifiable, the system can operate better - ideally, it is better to use tokens directly, which will be a clean closed-loop system. Compare the previous cryptocurrency trading system with the system for identifying tumors in X-ray. In the latter, you need to convince an insurance company that the X-ray model is valuable, and bargain on the issue of how much value it has, and then trust a small group of people to verify the success or failure of the model. This is not to say that a more active society of Digital Native content will not emerge. Like the recommendation system mentioned above, it may be very useful. If it is attached to the exhibition market, it is another case. The model runs on the chain, the system rewards tokens, and creates a clean closed loop again.
        Although it seems unclear now, I hope that the blockchain native projects will be expanded over time. First, the decentralized machine learning market can break through the data monopoly of the current technology giants. In the past 20 years, the giants have standardized and commercialized the value resources on the Internet, forming an exclusive data network and a strong network effect around the data. As a result, value creation shifts from data to algorithms. (with the standardization and commercialization of technology, we are approaching the end of the era of data monopoly network. The chart is from placeholder). Second, they have created the most powerful artificial intelligence system in the world.
        Attract the best data and models through direct economic incentives. Their strength has increased with the increase of multilateral network effects. With the commercialization of data network monopoly in the Web2.0 era, it seems that they may become the next winner. We may need a few more years, but the direction seems to be right. Third, as shown in the recommendation system case, the search will be reversed. Instead of people searching for products, products search for and win over people (this framework is attributed to Brad). Everyone may have their own curatorial market. According to the definition of individual relevance, the recommendation system and the algorithm model compete to recommend the most relevant content for them.
        Fourth, they will enable us to obtain powerful machine learning based services from companies like Google and Facebook without disclosing our privacy data. Fifth, machine learning will develop faster, because any engineer can access the open data market, not only engineers in large Web 2.0 companies. First of all, safe computing methods are still quite slow, and machine learning is still very expensive in computing. But on the other hand, as people have more interest in secure computing methods, everything is getting better. In the past 6 months, new methods for improving the performance of he, MPC and zkp have emerged.
        Cleaning up and formatting crowdsourced data is also challenging. We may see some combination of tools, standardization and small enterprises to solve this problem. Finally, it is ironic that the business model for creating a generic construct of such a system is not as clear as creating a single instance. This seems to be more applicable to many new encryption primitives, including the curatorial market. The combination of machine learning and blockchain incentives can create the most powerful machine intelligence in various applications. Over time, there are major technological challenges that can be addressed. However, their long-term potential is huge, and they may stand out from the current control of large Internet companies over data.
        They are also a bit scary & mdash& mdash; Because the system leads to its own existence, self reinforcement, and consumption of private data, it is almost impossible to stop and shut down. Is it a bit like calling the unprecedented powerful pagan god Moloch to create them? In any case, they are another breakthrough in how cryptocurrencies will slowly and then suddenly enter every industry.
Previous:Is the blockchain concept plate just a meteor shower
Next:No more

Related articles:

© 2005-2032 | Blockchain Circle & & All Rights Reserved    Sitemap1 Sitemap2 If there is infringement, please contact us at: