Machine learning models derived from data training based on the blockchain marketplace have the potential to create the world’s most influential AI. They combine two parts: private machine learning, which trains on private information without giving it away, with blockchain-based incentives that allow these systems to attract the best data and models, making them smarter. The end result is an open market where anyone can sell their data and keep it private, while developers can be incentivated to attract the best data for their algorithms.
Building these systems is very challenging and the required building blocks are being created, but from the current simple initial version it seems possible. I believe these markets will move us from the current Web 2.0 era to the Web3.0 era of open competition for code and algorithms, both of which can benefit directly.
The origin of
The idea came out of discussions in 2015 with Richard from the Numerai Foundation. Numerai is a hedge fund that sends cryptocurrency market data to any data scientist who wants to model the stock market. Numerai submits the best models to a “metamodel”, trades the metamodel and pays the data scientists who build the platform.
Having data scientists compete seems like a pretty good idea. So this led to more thinking: can we build a completely decentralized version of the system that can be applied to any scenario? I know the answer is yes.
create
For example, we are trying to set up a completely decentralized system to trade digital currencies on a decentralized exchange. This is one of many potential creation approaches.
Data The data provider owns the data and is used by the model builder.
Modeling The modeler chooses what data to use and creates the model. The training is performed using secure computations, which also allows the model to be trained without compromising the underlying data. The model needs to have different weights.
Establishing a metamodel A metamodel is created based on an algorithm that takes into account the weight of each model.
Creating a metamodel is optional – you can imagine many models that are not used in conjunction with the metamodel.
Using a metamodel Smart contracts use a metamodel and trade on a chain through a decentralized trading mechanism.
Distribute gains/Losses Over a period of time, a trade generates gains or losses. This profit or loss is distributed to the metamodel’s contributors based on how much they contribute. Those models that make negative contributions will have some or all of their collateral taken away.
Validation computations are centralized for each step, but validation and challenges can use decentralized systems like Truebit or secure multiple computations.
Storage data and models are stored on nodes such as IPFS or in multi-role computing networks because on-chain storage is too expensive.
What drives such a system?
Incentives to attract the world’s best data Incentives to attract data are the most important part of the system, because data is the limiting factor for machine learning. Similarly, Bitcoin has built the world’s most powerful network of computing power through open incentives, and the right data-driven architecture will attract the best data in the world to apply for you. And it is almost impossible to ban data from thousands or millions of sources.
- Competition between codes creates open competition between models/code, which has never happened before. Publish thousands of competing news delivery algorithms on decentralized Facebook.
- Providers that reward transparent data and models can see that they are getting fair benefits associated with submitting tasks because all calculations are verifiable, which makes people more willing to participate in such projects.
- Automation creates an automated and trust-free closed-loop by operating on-chain and taking value directly from tokens.
- Network Effects Multifaceted networks are influenced by users, data providers, and data experts, which also makes the system self-reinforcing. The better the system performs, the more capital it attracts, the more potential for returns, the more data providers and data experts it attracts, the smarter the system becomes, the more capital it attracts, a virtuous circle.
privacy
In addition to the points mentioned above, one of the main features is privacy. It allows users to 1) submit data that is too private to share and 2) prevent the economic value and model of the data from being corrupted. If unencrypted data is made public, data and models can be copied for free and used by others who make no contribution (the “free rider” problem)
Part of the solution is to make data sales private. Even if the buyer chooses to resell or release the data, its value diminishes over time. However, this approach limits short-term use cases and still presents typical privacy concerns. Therefore, a more complex but effective solution is to use some kind of secure calculation method.
Safety calculation
The secure computing approach allows the model to be trained without compromising the data itself. At present, there are three methods of secure computing in use and research: homomorphic encryption (HE), multi-party secure computing (MPC) and zero-knowledge proof (ZKPs). Multi-party computing is by far the most widely used algorithm for private machine learning, because homomorphic encryption is too slow and it is not obvious how to incorporate zero-knowledge proofs into machine learning. Secure computing method is the frontier of computer science research. They are often exponentially slower than ordinary calculations, reflecting the system’s bottlenecks, but have improved over the years.
Ultimate recommendation System
To describe the potential of personal machine learning, consider an app called the Ultimate Recommendation System. It can see everything you’re doing through your device: your browsing history, anything you’re doing on your app, pictures inside your phone, location data, spending history, wearable sensors, messaging content, the camera in your home, the camera on your AR eye, and so on. It then suggests the next website you should visit, the article you should read, the music you should listen to or the product you should buy.
This recommendation system can be very useful. Google, Facebook or any other existing database probably doesn’t have a system like this because it has the biggest vertical view of you, and the system learns from your private, non-leakable information. Before and said to the case of digital currency trading system, it will by focusing on different areas of the model (for example: website recommendation, music) to operate, compete to get users to encrypt data access and as users are recommended, perhaps even because users have contributed data and to focus and recommend things like they pay. Google’s joint learning and Apple’s differential privacy are a step in this direction of private machine learning, but still require trust, don’t allow users to verify security directly, and keep data private.
What is likely to come first?
I can’t be very precise about what construction is best, but I have a few ideas. One of the criteria I use to evaluate blockchain solutions is this: the more blockchain native, the better the range of research — from physical native to digital native to blockchain native. The less blockchain native, the more third party involvement is required, making for increased complexity and reduced ease of use with other systems as building blocks.
Here, I would argue that if the value creation in the system is qualified, it means that the system is more likely to work successfully — directly in fiat currency terms, and a better choice is tokens. So you have a pure, closed-loop system. A comparison can be made between previous cryptocurrency trading systems and X-ray tumor recognition systems. For the latter, you need to convince the insurance company that the X-ray model is worth it, and negotiate how valuable it is, and then trust a small number of current people to make the model a serious success/failure.
This is not to say that positive sum situations will not occur when societies use digital native systems. A recommendation system like the one mentioned earlier can also be very useful. If connected to digital markets, there is another use case where the model can run code on the chain, and the system rewards generation (for digital market cases), which creates a pure closed loop. It may not seem obvious right now, but I expect blockchain-based native tasks to expand over time.
impact
First, a decentralized machine learning market could remove the monopoly on data held by existing technology giants. They have standardized and commoditized the Internet’s main source of value creation over the past 20 years: proprietary data networks and the powerful network effects that surround them. Therefore, value creation begins to shift from data to algorithms.
In the cycle of standardization and commercialization of technology, we are now at the end of the value monopoly of the Internet age. In other words, they created an intuitive business model for AI.
Second, the decentralized machine learning market has created the world’s most powerful AI systems, attracting the world’s best data and models through direct financial incentives. Their strength grows with the efficiency of multi-party networks. As data monopolies become commercialised in the Internet 2.0 era, they look like candidates for the next convergence point. Maybe we’re still a few years away, but it’s in the right direction.
Third, as the recommendation system shows, search is inverted. People don’t search for products, but products search and compete to serve people. Everyone may have their own favorite market, recommendation system can show the most relevant content, and those content is very relevant to the personal definition.
Fourth, a decentralized machine learning market allows us to achieve the same revenue as Google and Facebook without having to give away our data.
Fifth, machine learning can advance more rapidly because any engineer can access data in the open market, not just a small group of engineers at a few big companies in the Web2.0 era.
challenge
First, secure computing models are now slow and machine learning has become computationally expensive. On the other hand, the performance of secure computing is also gradually improving. I have also seen some solutions that have achieved significant performance improvements for the HE,MPC and ZKP over the past 6 months. It is difficult to compute specific data or model values and feed them into the metamodel. Cleaning up and formatting crowded data can be challenging. We’d like to see tools, standardization and small businesses come together to solve this problem. Finally, the business model for creating a generalized construct of such a system is less clear than creating a single instance. This is true for many new crypto things, including select markets.
Giiso Technology, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.
conclusion
The combination of private machine learning and blockchain incentives can create the strongest machine intelligence in a wide range of different applications. But there are still several very serious technical challenges. Their long-term potential is huge and will change the way existing big Internet companies own data. It’s also a bit scary, because such systems can exist, self-enhance, consume private data, and are almost impossible to shut down, making me wonder if creating them would summon a more powerful Moloch. Either way, this is another case of how cryptocurrencies can slowly develop and then suddenly enter any industry.
Above, Chen Ruchu’s humble opinion! I’m sorry if I offended you.
For more information about ai robots, please visit Giiso Zhisou :http://www.giiso.com/. Thank you