By investigating six use cases and future applications, we demonstrate how AI entities can utilize the capabilities of blockchain for important purposes including, but not limited to, deep learning, Internet of things (IoT) and Monte Carlo analysis. We also explore the possibility of storing externally trained AI agents on such a medium and utilize them via pay per use. Finally, we describe an already trained neural network employed to recover the relevant physical variables, both in quantum and classical systems.
3.1. The Integrity and Validity of Information
Blockchain as a data and framework store presents a number of advantages over the Internet or over internal stores. By way of two exemplary challenges to the AI world, we present how blockchain can address these in novel ways.
One of the biggest challenges in data science today is the collection of a proper dataset, which can be utilized for training a neural network. The pluralism of data over the Internet is enormous, but the quality is minimal due to the habit of people to post inaccurate things, mainly, because there is no control. A characteristic paradigm is the “fake news” explosion in recent years, which tends to propagate faster than well documented and verified news [25
]. Internet giants like Facebook and Google have tried to tackle the problem via several computational methods, but even though there seems to be a sufficient theoretical basis for separating “signal” from “noise” [26
], the problem still thrives as of today.
A second challenge is adversarial interference with the processing. Tesla’s autopilot was shown to be vulnerable to remote root privilege attacks that could control the steering system and disturb the “autowipers” function [27
]. By introducing false information in the physical world such as minor changes in the road, it was possible to mislead the car into the opposite lane. The consequential risks of such vulnerability include, but are not limited to, human injuries and death. Many other examples abound. Blockchains can address these issues in a comprehensive way through integrity, security, triple entry and provenance.
Data as fact integrity
: The cryptographic inventions of digital signatures and hashes have led to a general technique for making data reliable within the context and limitations of the technical means, a characteristic called integrity. In practice this means that we can state with (cryptographic) certainty that a piece of data existed no later than a particular time, and that it remains untampered with. These cryptographic techniques need some software to deliver results. Timestamping [28
] involves taking the hash of a document and placing it in a timed sequence of hashes that is kept alive essentially without limit on time. Each new document’s hash is placed in a block, which is then hashed, along with a hash of the last block and the current time. As the cryptographic hash is essentially unforgeable without the actual block, this ensures both the inclusion of the new document(s) and the proof that the last block, and by induction all previous blocks and included documents, are securely timestamped. The reliability of the stamp of time is the reliability of the recording of the time in each block, and the space between the blocks.
Facts by people, securely: Digital signing takes the evidence of a hash one step further by indicating who it was that made that stamp. Digital signatures are made by a private key, and verified by a public key, which latter also takes the form of an identifier for the private key called a pseudonym. This security model is essential for a blockchain as it ensures that only the proper pseudonymous agent, as holder of the private key, can make new transactions. Money is perhaps the most harshly attacked activity of humanity after wars, and therefore can only survive if protected by strong security. The cryptographic security model of pseudonymous digital signing used in blockchains is battle-hardened and is available for free for all other applications beyond transfers of value. This is no trivial benefit as the Internet has quite poor security models, and big Internet applications such as online banking and autonomous vehicles generally have trouble deploying robust security to users. Injection of information from unknown sources is rampant, and simply adding data stamping and signing as used in blockchain makes the attacker’s job harder.
Facts as shared knowledge:
A technique known as triple entry accounting [29
] adds a further advantage captured by the aphorism “I know that what you see is what I see.” Triple entry takes the above integrity techniques and makes records such as offers and acceptances, payments, receipts and invoices both shared and reliably the same to all relevant parties, which allows software to work with reliable raw data as facts produced by other parties; triple entry accounting does for trading groups what double entry accounting did for the firm. Independence from weak data, whether summarized, prepared, or sanitized, results in the elimination of diverging data sets and unreliable outcomes. For example, clearing and settlement in financial trading is highly simplified if the data is already guaranteed to be the same for all.
Blockchains go further and incorporate a public database that ensures everyone has access to the same data, and some parties are financially motivated to keep that database alive. This ability to always find the data comes at the cost of privacy—whatever is posted to the blockchain as a document is readable by all. There is some promise of more exotic cryptography and software techniques to allow posting and recovery of private data into a public store, but these techniques remain experimental, and the bar of confidentiality or privacy is typically a high one.
Knowledge as truth: What remains is the provenance of the data at the time of posting. The blockchain supports two easy controls, and one hard control. Firstly, if the data is a financial transaction on a blockchain, in an asset mediated by that very blockchain, then the transaction record can support its own provenance, gained in part that someone went to the effort of moving money, and in other part that it cost a small fee. Secondly, the use of the pseudonymous digital signatures provides a minimal form of identity system: A document’s utility and provenance can be analyzed within the context of all the documents posted by the same agent. If Alice generally posts good documents, then the next is likely to be good; if Bob posts fake news then people should expect more of the same. Pay on demand is discussed in the next section.
Consider two trivial statements, “this statement is true” and the equally light “this statement is false”. Both can as easily be posted, but only one is reliable. Software can guarantee both statements were made at a time, but cannot guarantee the content is reliable or even meaningful.
Then, to encourage statements that may be relied upon by others requires more: Posters need to be incentivized to post useful and reliable statements, and to not post useless and unreliable ones. Due to the pseudonymous nature of blockchain, posting stake or gamification is suggested as a control however these methods limit participation through the cost of capital and time, and leave aside the question of how to punish [24
]. A more serious feedback control on bad participation would be a due process to also incentivize agents to not post unreliable data. The process itself would also need to pass the same test of reliability as the statements delivered.
Such a due process is typically called Public Key Infrastructure (PKI). The more common Internet secure browsing form organizes a certification authority to make signed statements, called certificates. Its due process is described in documents such as a certificate practice statement, which are reviewed and approved by browsers and other relying parties. Reliance based on commercial authorities and their statements is typically only strong enough for relatively weak statements because it lacks an incentive model to properly handle the liability for bad data [30
]. CAcert has extended the concept to cover a wide range of stronger statements through a cooperative form that includes arbitration to allocate liability in the case of bad data [31
Blockchains are therefore not only ideal storage for the data of deep learning, they include much data worth analyzing, and they are ideal storage for the trained frameworks themselves. In time we expect the discrimination between good and bad data to become easier based on pseudonyms and incentive models.
3.2. Programs Stored on Chain and Composed Within Transactions—Pay Per Use
As above, a blockchain forms a novel method to store information on a public space, via a payment procedure [32
]. As well as static information such as literature or news, we could also store the code for programs in much the same was as Github [33
], and indeed, the underlying git system is very like a blockchain in many respects. These programs can be read freely as they are part of the immutable data of the chain. Each transaction that posts data on the blockchain costs money and thus it is uneconomical and non-incentivized to continue posting programs unless there is at least a minimum revenue possibility.
A collection of on-chain applications can be made browsable via a portal such as Agora [34
], forming a new channel for distribution of software (Figure 4
). This lays the foundation for a long held ideal of programmers, being an independent marketplace where the developers can be paid for their work without any intermediates [35
]. The space is fairly new at the moment of writing but it does not lack for novelty and innovation, including applications in art, music, money and weather. Other applications that could fit include IoT sensors over power grids, security systems or transport networks.
By employing the OP_RETURN op_code instruction of Bitcoin Script [7
], a new world of application utilization emerges. Transactions can refer to and run other programs on a “pay per use” basis, allowing for programmers to ‘compose’ larger programs out of many smaller ones. An example of pay per use is found in Moneybutton [36
]. With such tools it is possible to construct a ‘Metanet’ being an immutable version of Internet as we now experience it.
3.4. Artificial Intelligence Agents Trained Via Submitted Blockchain Data and Operated on Chain.
An artificial intelligence agent (AIA) goes one step further than a trained framework by utilizing the new data in the user’s request to advance the neural network forward by a new epoch; in other words, it learns as it works. Fitness is determined when minimal or no other changes are required with any future submitted data. Let us consider an example extended from above. In the recent years, a plethora of authored code in various programming languages has been stored in repositories such as GitHub [33
]. Complex algorithmic programming is a time consuming and costly task; a programmer requires a high intellect and years of education, and complex tasks often require many months of collaborating work between several parties.
Blockchains can assist in two ways. Firstly, as above, a blockchain can store the code. Secondly, an AIA, encoded on the blockchain, can assist the programmer in many ways: Conversion of code from one language to another, searching for algorithms that match patterns, conformance of requirements or documentation to code and eventually in authoring new algorithms. Using deep learning techniques and big data mining from existing code repositories, this AIA would present a reliable, secure and disruptive technology.
AIAs are described in computer science as abstract entities that are able to monitor and evaluate certain parameters through various input sources (i.e., IoT sensors, I/O raw data, databases, ontologies, etc.) towards achieving a rational goal [39
]. Their basic role is usually described as that of an actuator, with the simplest implementation of an AIA to be derived from a reflex machine, such as a thermostat, but they can vary from very simple to extremely complex. There are four architectures of AIA to be considered [39
]: (a) Logic-based agents (decision of action is derived via logical inference), (b) reactive agents (decision is based in some form of direct mapping from situation to action), (c) belief–desire–intention agents (decision depends upon the manipulation of data structures) and (d) layered architectures (decision is realized via various software layers, each depended on its environment at different levels of abstraction). There are also five classes of AIA to be considered [40
]: (a) Simple reflex agents, (b) model-based reflex agents, (c) goal-based agents, (d) utility-based agents and (e) learning agents.
All the above AIAs learn from the submission of fresh data, e.g., code from programmers from anywhere in the world, who will be incentivized to post their work on the blockchain, either for gaining revenue, or under a specific license (i.e., Opensource, MIT, etc.). Moreover, a new opportunity emerges through the concept of AIA paying another AIA for services or paying a sensor for data. An example is WeatherSV (Figure 6
), which offers weather prediction to users in their selected territory by utilizing the data collected by a set of global IoT sensors. The service can be activated for a cost of $
5 AUD and delivers hourly reports for about 123 days, based on current fees of Bitcoin SV [41
The notion of live data feed and immutable storage can form a basis for implementing several other applications such as decentralized logistics, as shown in the work by Christodoulou et al. [42
], also as a supply chain for manufacturing, the agricultural sector or even a modern city. A Smart City [43
] that uses different types of electronic IoT sensors to collect data and then uses the data to manage assets and resources efficiently can use the blockchain as an immutable record ledger both for integrity, for deep learning and also for historical purposes. For example, Zweispace in Japan now stores the national earthquake sensor data on the blockchain [44
By utilizing the proper AIA or CNN with the data provided on chain, we can have productive outcomes towards the implementation of a far more economic and robust Smart City economy. A similar opportunity applies wherever there are vast amounts of both data and users learning from that data: Traffic control and other problems in transportation and supply chain, education and health. Additional applications of AIAs on chain can facilitate analysis of financial markets, DNA for rare genetic disease detection, high definition imaging of stellar bodies for possible collision detection, auditing and protection of network against attacks and more.
A collaboration of AIAs as a swarm is also applicable via payments in the form “Machine paying Machine” in order to achieve solutions to more complex tasks.
3.5. Proof of Work Via dSHA256 as a Source of Randomness and Monte Carlo Method Via ASICs
The concept of proof of work (PoW) as introduced in Bitcoin is a reward mechanism to the solvers of a random puzzle. A hash puzzle is a set of mathematical problems, which are solved by creating a hash that conforms to a specific requirement, being firstly a hash over a new proposed block. Secondly, in the block’s header, an extra value called a ‘nonce’ or ‘number-once-used’ is cycled repetitively to produce a trial hash value with a large number of leading zeros.
Solving the puzzle is competitive and thus computationally difficult. Unless the cryptographic hash function used for calculating the block hashes is broken, the only fruitful strategy is to try different nonces until a solution is found [45
]. Bitcoin uses the SHA-256 hash function [46
], which is a leading standard for hashes.
The fastest participant to find and propagate a winning solution is rewarded. Bitcoin also includes two feedback loops that vary over time. Firstly, the difficulty, or the minimum threshold of number of zeros, is varied every two weeks to keep the expected time to solve around 10 min. Secondly, the reward paid for solving the puzzle halves every four years.
At the time of writing, the reward stands at 12.5 Bitcoins [47
] and the Bitcoin Hashrate is estimated on average at 53.85 Eh/s (SHA-256) [48
]. That gives us 53.85 × 1018
random numbers per second, in effect, making the miners pseudo-random number generators (PRNGs). The result of this is the generation of an impressively large number of random numbers, for every block. A well-known computational method that is capable of providing solutions in Non-deterministic Polynomial-time – Hard (NP-hard) and Non-deterministic Polynomial-time – Complete (NP-complete) problems via utilizing the random numbers that a SHA256 miner can produce is the Monte Carlo method. Monte Carlo is a category of computational algorithms, which is based on continuous and repetitive random sampling in order to solve complex problems. The underlying concept is to use random solutions to solve problems that can be deterministic in nature. The method is often used in physical and mathematical cases and is very useful when it is difficult or impossible to use other approaches. The Monte Carlo method is used in three categories of problems: Optimization, numerical integration and guessing results from a probability distribution [49
As a side-effect of PoW, blockchains can extend their activity to solving massive Monte Carlo problems. Blockchains are in effect the biggest PRNG in the world right now and probably the fastest PRNG swarm that could solve literally any computationally hard problem via utilizing the Monte Carlo method.
3.6. Solving Physical Problems via CNNs and Simulation of Quantum Computing
There are numerous studies demonstrating the abilities of a neural network. For instance, a deep neural network is able to learn through training and produce fairly accurate predictive results correlated to the dataset they trained on, while recurrent neural networks are being used towards the deterministic analysis of speech recognition, down to video prediction [50
]. Neural networks can be trained offline and then can be stored on the blockchain and parsed via pay per use. On another note, because of the fact that users will be using the medium to submit data the entity can evolve to a higher scale and store a more advanced version of itself for later use.
One of the most difficult tasks in neural network mechanics is to understand how they function and how they are able to extract results. They are often used as “black boxes” and consequently our perception of the mechanics of them are limited. Several studies have tried to analyze the inner-works of CNNs but we believe the most prominent way to answer this question is the simulation of physical concepts and the solution analysis.
What we propose is not new; several studies have tried to employ a CNN towards simulation of a “human-like problem analysis” and the results were quite impressive. We now know that a Region-based Convolutional Neural Network (R-CNN) can be trained recursively to analyze problems via data pretty much like a human brain can. Moreover, the outcome is extremely accurate to the expected results. The materialization graph of SciNet, a CNN that represents the aforementioned process is displayed below (Figure 7
In the aforementioned work and by utilizing the same concept it is shown that, based only on (simulated) experimental data and without being given any assumptions about quantum theory, SciNet recovers a faithful representation of the state of small quantum systems and can make accurate predictions (Figure 8
The implementation of neural networks based on similar technologies when implemented on the blockchain can dramatically amplify their computation abilities via the utilization of a PRNG engine provided by the PoW procedure as per solving complex problems. Furthermore, such a computational entity can provide solutions to many problems that are now impossible to approach (e.g., deterministic computing and chaotic system analysis).