AI And Decentralization
By Farshad Kheiri, PhD, MSc. on ALTCOIN MAGAZINE
1) Introduction
The fourth Industrial Revolution is known as the digital revolution where technology is conquering every aspect of all active industries and human life. Artificial Intelligence (AI) and decentralization will play a dominant role in this era. In this article, AI and decentralization will be explored to answer how they co-exist to push forward to the next revolution.
2) What Is AI?
AI is defined as a computer program that completes a task that if it is done by a human, requires intelligence (Marvin Minsky and John McCarthy). Machine learning works as a toolbox for AI where mathematical algorithms are used to discover patterns from the data. This data could be gathered over time (known as historical data.). With the right hypothesis and algorithms, machine learning algorithms are created using historical data. The process is called training a model. In this case, these models can help to predict the future.
For example, if there is a correlation between the amount of ice cream cones sold in a beach city and the number of deaths, one can naively conclude that the number of deaths could rise if the amount of ice cream cones sold rises. However, applying the same logic to a city far from the beach may not give us a desirable result! A more knowledgeable individual (expert) may simply interpret that the rise in ice cream cones sold is a response to warmer, which results in more people going to the beach. Going to the beach and swimming may also create a higher risk of drowning and other health risk factors. This may explain why the same regression model may not work for a city that is not next to the ocean.
Over the past 12 years, I have observed several people interpret data incorrectly. Nowadays, there are several tools that enable people to build models easily. (Just find some sample code on the Internet and place it in your data!). Obtaining incorrect results has become very frequent and some blame data science as a hoax! Yes, a Chief Risk Officer from a famous bank, once on LinkedIn was blaming that the entire data science field is BS! I had the privilege to work with this gentleman and on several occasions, I noticed that he had a wrong interpretation.
Correlation does not imply Causation! Building more complex models is not always the best solution! Having more data is not necessarily good if the integrity of the data is not checked! These simple rules and others have been ignored by many intelligent people over time. Building a model can’t be just a rigorous mathematical exercise, building a good model needs intuition, intelligence, and good data. I remember one of my mentors, once called building a model “an artistic task.”
In this article, I am not going to go through the details and examples of AI, but just to be clear Artificial Intelligence, Machine Learning and Deep Learning are all related to each other as shown in Figure 1.
3) What Is Decentralization?
Decentralization developed from the word “centralization,” as a term to decentralize the government (Maurice Block). Over the past 200 years, it has attracted a lot of intellectual leaders and scientists. Decentralization entered technology discussions in an article by Satoshi, which led to the creation of Bitcoin — a digital currency that does not require bank verifications for the transaction. Blockchain technology has been around since 2008 as a decentralized database. All Bitcoin records, known as blocks of data, are connected through cryptography and consensus of the individuals who form the distributed blockchain. Eventually, other technologists built on the blockchain by allowing users to run code on the Blockchain decentralized database. This uses what is called smart contracts, a series of codes that can’t be tampered and which are revealed to all users. For example, if User A fulfills a series of terms, User B gives him or her a certain amount of tokens. These tokens could have some economic value outside of this environment, or just be necessary to run anything in the environment.
Over the past 12 years, scientists and engineers have developed different strategies and techniques to address a lot of the rising issues. This includes how much a token should be worth, how should a right action be rewarded, while a malicious action is identified and punished. How to stop attacks? How to prevent tampering with the ledgers? All these resulted in a massive growth in Game Theory, Cryptography and Token Economy strategies. In Figure 2 below, you can see how decentralization, blockchain, and Dfinity as one specific blockchain technology, relate to each other.
4) Areas These Two Can Benefit From Each Other?
So far, we have a scientific field that has developed based on data, and another branch that developed as a trustworthy database which can identify the involving sides in a transaction. Decentralization will enable us to address three important issues:
a) Ownership
The data are written into the blockchain, is being managed and maintained by the community, as long as the majority decide to honor the ownership standards,
b) Privacy
If the user has ownership, she or he can decide what to share or not to share and this will enable privacy of the data.
c) Trust
If the user knows that the ownership of the data won’t be challenged or revoked and she/he can decide where and when to share the data, she/he will feel trust to keep data in the database and even add more information.
In the future, we will explore how the three aforementioned criteria, can be applied to AI. We will write separate articles on the following examples in the future.
a) Federated Learning, which ensures that user privacy and data are enforced. Federated Learning has been used to build a model on a user device. The model can be shared with other models to improve the model, meanwhile, the data never leaves the user’s device. Federated Learning implementation can occur on a decentralized platform and through smart contracts, the user can decide when and where to share a model with others. One of the main applications of this algorithm will be healthcare, where sharing data is restricted by HIPAA. However, if the models encrypted and the decryption keys owned by the model owner. They can decide with who to share the model. Also, based on the characteristics of the data, through a decentralized program, the best candidate models to be ensembled, can be identified.
b) Pattern Discovery involves using continuous learning on the data which has a time-stamp. Pattern Discovery will make learning processes more accessible and identifying patterns easier. The trained model can be updated as new data comes in. Also, the user can decide what data to share or not to share. For example, if a user is buying a gift for someone (this may be a 1-time purchase), this data does not need to be included in the buying history of the user. The purchase history of the user can be a strong tool to predict and recommend the next purchase. However, many users feel uncomfortable to share all the transactions and sometimes, the purchase is one thing, e.g. a gift for a friend. Through a decentralized platform that is being controlled by the user, she/he can decide what to be shared. The time-stamped data makes sequence pattern discovery easier and the models can be updated as the data comes in.
c) Label Visibility is one of the challenges that currently exist with some AI models is a lack of trust in how the models are consuming the user data and feedback. For example, if a user flags an article as not trustworthy (fake), will the platform continue to promote it, especially if it may be clickbait? Through recording user feedback, on a decentralized platform, and training an AI model with the data, users may trust that their feedback is accurately recorded.
d) Decentralized Partially Observed Markov Decision Process is one of the most interesting AI algorithms, which has been designed for a decentralized platform. Using the Markov Decision Process to aggregate learning from decentralized agents. Autonomous robots and cars can be some of the main users of these algorithms. Using a local processor in the car/robot, the data can stay local and the inductions can be shared with other agents.
e) Data Integrity brings to mind the famous sentence ”garbage in- garbage out,” which can be applied to blockchain technology. Most decentralized platforms are attempting to check the integrity of their data through some human intelligence, named validators. This might be do-able on a small platform, but for a large platform and special data, the individual validator may be expensive and not scalable. This is where AI can play a great role as a real-time, scalable validator. Medical research has had serious challenges to access data and more importantly access authentic data. Many of the data integrity validation techniques can be modeled by AI and by integrating the informed user feedback through continuous learning, the models can be updated regularly.
5) Challenges
Blockchain technology as a way to implement decentralization has its own challenges. It is not fast enough and faces difficulties in implementation. Therefore, the architect of these platforms must develop a structure in which fast speed does not hurt the functionality of the platform. Also, blockchain can’t handle massive data. Again, the architect of the platform must decide what should be placed on the chain to avoid unnecessary burdens on the blockchain.
Saying all these and being skeptical as scientists, we still have tremendous excitement about decentralization, blockchain technologies, and we think that AI will benefit highly from it. In the era of the new Internet, the user won’t be targeted by a marketing agency that attempts to manipulate their purchase decisions. In the era of the new Internet, the user’s data will be controlled by the user and she/he can decide how this data is used. Rather than being a target, they can be intelligent shoppers and users of the most fitted service to their needs.
