I’ve been working on a new series of papers on decentralized AI and decided to edit and series I published early last year.
The emerging field of decentralized artificial intelligence(AI) is becoming one of the most exciting technology trends of the last few months. A lot has been written about the potential value of the intersection of artificial intelligence(AI) and blockchain technologies and we, this year, we have even entire conferences dedicated to the subject of decentralized AI. However, I feel that a lot of the hype behind decentralized AI fails to highlight some of the key value propositions of the new technology movement that can make it one of the most foundational technology trends of this decade. If you believe in the idea that AI is going to become an increasingly influential factor in our daily lives, I believe decentralized AI will be an essential element to guide the impact that machine intelligence will have in future generations. Sounds dramatic? Let’s look at some of the economic dynamics behind decentralized AI to try to clarify our point.
These days, the notion of AI systems is intuitively linked to centralization. The first thing that comes to mind when we talk about AI are companies such as Amazon, Facebook or Google whose machine intelligence systems are becoming part of our daily lives. The increasingly rich data assets possessed by those companies have allowed them to capitalize first on the AI revolution and create an economic dynamic that is not always aligned with the end consumer. Even the technology and methodologies we used today for building AI systems assumes a centralization model at its core.
The lifecycle of a modern AI project assumes that you have a model and a massively large, high quality dataset that you can use to train it as well as a pool of data scientists that can constantly regularize and optimize the model in order to become more intelligent. In most AI scenarios, that entire cycle is performed by a single entity that has the resources to collect large datasets, create highly sophisticated AI models and run across expensive computing resources.
The irony of all this is that, when you look deeper, the economic incentives of the large party providing the AI models are not necessarily aligned with the value creation for consumers. From an economic standpoint, there can be many scenarios in which the ability of an AI agent to increase the value of the assets of its creator in the form or revenue, data or simple outcomes is not directly correlated with the ability of creating more value for consumers.
Centralized Intelligence vs. Federated Knowledge
The centralized nature of AI systems highly contrasts with the evolution of human intelligence. Knowledge exists completely scattered and federated across the world. Erudition is a novel goal in live but nobody can claim to posses all knowledge of a particular subject. Knowledge collaboration and federation is one of the key unique advantages that allow humans to evolve and dominate other species that were physically more powerful. And yet, AI remains increasingly centralized. In a world that is moving rapidly towards the creation of general AI and systems that can vastly surpass the level of intelligence of mankind, wouldn’t we want that knowledge and influence to be federated instead of controlled a few organizations?
The emergence of technologies such as mobile computing or internet of things(IOT) challenged the centralized notion of AI. Today, knowledge is constantly created in the edges and flows towards centralized hubs. The pendulum has to shift to a dynamic in which aspects such as the training, optimization, testing and knowledge creation of AI model becomes federated across many participants.
In order to decentralized AI models, we need to solve a few challenges:
a) The Privacy Problem: Can entities train a model without having to disclose their data.
b) The Influence Problem: Can third parties contribute to the behavior of knowledge of an AI model in a way that is quantitatively influential.
c) The Economic Problem: Can third parties be correctly incentivized to contribute to the knowledge and quality of an AI model.
d) The Transparency Problem: Can the activity of behavior of an AI model be transparently available to all parties without the need of trusting a centralized authority.
Centralized AI Today is Like Closed Source in the 1990s
Open source today is highly rewarded and the best and most efficient way to create software but that wasn’t always the case. For decades, large software companies preferred to embrace closed source delivery models in order to have an edge in terms of intellectual property(IP). Eventually, the economic dynamics proved that thousands of talented engineers regularly contributing to a project produce better code than a few engineers driven by corporate interests.
If we extrapolate the evolution of open source to the AI world, today we are somewhere in the 1990s in which the value creation of software was controlled and influenced by a few companies. What is worse, when comes to AI, we are not only talking about software or AI models, but also other expensive resources such as data science talent, data and computing power. In that world, decentralized AI is the new open source except that the impact in mankind can be order of magnitude more impactful to mankind.
Despite its somewhat obvious value proposition, the path to decentralized AI was plagued with very difficult technical challenges that made it completely impractical in real world applications. From the pure technological standpoint, many of these problems were considered unsolvable until recently. In the last few years, new technologies in the cryptography, digital currencies and AI space has come together to provide a solid foundation for the implementation of decentralized AI applications.
The Privacy Solution: Homomorphic Encryption
Mathematically, homomorphism is defined as “a mapping of a mathematical set (such as a group, ring, or vector space) into or onto another set or itself in such a way that the result obtained by applying the operations to elements of the first set is mapped onto the result obtained by applying the corresponding operations to their respective images in the second set”. Homomorphic encryption allows specific types of computations to be carried out on ciphertext which produces an encrypted result which is also in ciphertext. Its outcome is the result of operations performed on the plaintext. For instance, one person could add two encrypted numbers and then another person could decrypt the result, without either of them being able to find the value of the individual numbers.
Homomorphic encryption can be considered one of the greatest breakthroughs in the cryptography space in the last decade. In the context of decentralized AI, homomorphic encryption enables participants in an AI application to contribute data to training of a model in a way that it remains encrypted to the other parties.
The Economic Solution: Blockchains
Blockchains provide the fundamental runtime and protocols to enable true decentralized AI applications. The first generation of dencetralized AI applications are leveraging concepts such as smart contracts or DApps to model the interactions between different endpoints in an AI application.
Digital tokens are also a relevant concept in decentralized AI application as it represents the main mechanism to compensate data scientists for their contributions to a model. Digital tokens also provide an economic channel to guide and influence the behavior of models in a way that benefit all the interested parties.
The Influence Solution: Federated Learning
Federated learning is a new learning architecture for AI systems that operate in highly distributed topologies such as mobile or internet of things(IOT) systems. Initially proposed by Google research labs, federated learning represents an alternative to centralized AI training in which a shared global model is trained under the coordination of a central server, from a federation of participating devices. In that model, the different devices can contribute to the training and knowledge of the model while keeping most of the data in the device.
Its not hard to envision why federated learning is foundational to decentralized AI platforms. Using federated learning, multiple participants in an AI applications can independently train or optimize an AI model without having to trust each other or a centralized authority.
With economic and technical factors colliding, the trends in favor of the decentralization of AI seems more viable than ever before. However, the applications of decentralized AI remained incredibly constrained. I will explore this thesis in more details in a new series of articles.