Potential Of Collective Learning Using Differential Privacy

Potential Of Collective Learning Using Differential Privacy
Potential Of Collective Learning Using Differential Privacy

Imagine a hospital in rural Texas can have the same AI capabilities as a research hospital in New York City running cutting edge AI to help with the diagnosis while preserving patient privacy. This is the new world of collective learning. Collective learning is an application of deep learning algorithms that can change how we view data sharing and privacy.

During our Covid-19 pandemic, hospitals in rural China struggled with similar patient symptoms and issues as hospitals in a major city in Africa or Europe. If we have an AI diagnosis tool with collective learning to help clinicians and researchers worldwide, we can discover treatments and make a diagnosis that much faster.

Ever since Humayun Shiekh left Deepmind, he’s been preoccupied with the vision of the future. It’s a future when Blockchain enables decentralized AI. It’s a future where secure connections and democratized data sharing can preserve individual privacy.

Recently, his company, Fetch.ai, a Cambridge-based artificial intelligence lab, achieved a new milestone, demonstrating the potential of collective learning using differential privacy by the launch of its CoLearn Network.

We all know that knowledge is different than data. Using sophisticated artificial intelligence or human intelligence, data can be transformed into knowledge that can help us make better decisions.

The central issue surrounding data in the world of artificial intelligence is privacy. In healthcare, more than in any other area, privacy is a big concern. It’s the reason that currently, there aren’t more proponents of decentralized AI initiatives.

How can we share data as knowledge without losing the privacy of the data?

Shiekh says, “How do you take 10 people who might have different datasets to share and learn from that data without compromising privacy.”

In the current culture of Privacy, AI-enabled applications are anonymizing data to preserve Privacy. But, this type of implementation has its limitations.

Differential Privacy, is the idea of publicly sharing information about a dataset by describing the patterns of groups within the data while withholding information about the individuals in the dataset. It preserves Privacy while going one step further in the anonymizing process.

Shiekh says, “Previously, if you wanted to anonymize data, there was less of a problem with privacy, but, if you aren’t anonymizing it, then you need a different privacy solution. By incorporating blockchain, you’ve created a near perfect solution to the problem.”

With differential privacy, democratizing AI finally feels like an achievable vision instead of a far-off dream. This is the reason why Fetch.ai’s demonstration of collective learning based on differential privacy is a groundbreaking step forward for the AI industry.

Shiekh says, “Our CoLearn Network, uses differential privacy, which means that the whole system enables you to own your data while training models and participants outside of your network to learn from it.”

Fetch.ai’s CoLearn Network consists of hospitals and private practices using machine learning models to train X-ray image data from nodes in a distributed network between LA and London. The model correctly identified COVID-19 patients from pneumonia or other causes with 97% accuracy.

Think smart cities where each agent, whether it’s a self-driving vehicle, a traffic light, a parking spot, and people are all connected to achieve an optimal flow. In smart manufacturing plants, where parts, equipment, and production personnel interact with ease and transparency, managing change becomes a natural part of the process.

Smart hospitals connect seamlessly with other hospitals worldwide to share valuable knowledge to aid in patient diagnosis. It doesn’t matter how rare the symptoms, chances are, a doctor in a different part of the world has seen this before, even if this case hasn’t been published in a research journal.

Shiekh says, “One use case of our technology is its ability to be entirely borderless.  You can have hospitals from China to Cambridge working collaboratively to learn and improve the model so that when a hospital in Africa, who may not have the resources to implement an AI-learning system, can share their x-rays and get a result without having been a part of the training dataset. They don’t have to be concerned with the data and they can still get near perfectly accurate results. That is quite a powerful application because it solves that problem of silos. It also solves the problem of privacy.”

These are all use cases for decentralized AI where we operate as agents registered on a Blockchain network. Connectivity does not just mean connecting infrastructures, but connectivity also means that agents can share intelligence to improve their capabilities.

Each interaction on this vast network where we work alongside AI to improve our decision process also enhances the entire network.

Shiekh says, “Let’s start to put new models into traditional business models. What are we looking at? The key factors are, it’s many businesses, materials, applications that are all multi-stakeholders. That’s how we have to start looking at blockchain, one of the biggest multi-stakeholder deployments and applications out there, which is  building the connectivity for a digital economy… There’s also a lot of integration, a lot of multitasking. Using a multi-agent system, we can accomplish that.”

Can you envision the next generation of websites?

We are already in a world with more autonomous actions. With the help of AI-enabled systems, we are upgrading our intelligence with Big Data, bandwidth to process that data, and coming up with decisions that we could’ve never made without the help of technology.

In the new world, with the help of AI, we are agents connecting (literally) to each other across the globe. Each agent can have their own characteristics and their own autonomy.

But, the idea is that every agent operates within the limits of this world harmoniously. What does that mean in a Smart City? What does that mean in a factory? What does that mean in a company?

An agent of a smart vehicle registered in a Smart City will need to know how many free parking spots are available in the downtown area and which one the best one for the vehicle, given the current route that the vehicle has taken.

Shiekh says, “Now, we are evolving from websites to more autonomous actions. But, where’s the infrastructure? There’s no infrastructure. How do you get the agents to connect with each other and then to go one step further and  transact with one another. How do they learn from each other? This is collective learning. We built the Fetch.ai technology stack to solve for this. We are building the framework which enables you to launch an agent and do useful work through the agent. The infrastructures can also learn from split datasets, which either these agents could produce, represent or use underlying data.”

Blockchain offers the transaction transparency, security, and consistency needed to achieve differential privacy and decentralized AI truly. Without Blockchain, there’s never really “proof” that the pipeline is entirely fair or secure.

Shiekh says, “Without blockchain, there’s no immutable proof that anyone cheated. That is the biggest problem. You need to collaborate without releasing the data.”

One of the most significant oppositions of decentralized AI seems to be surrounding the profit model. If you can’t make money from proprietary intelligence (AI), how do you make money to sustain a business model? With differential privacy, here’s another possibility.

How about shifting the profit model from intelligence to data instead?

For tasks where companies already developed excellent proprietary algorithms, but where data can still be consistently improved, why not leverage decentralized AI to establish a payment model for the data? While each participant retains the data’s control and privacy, each participant can share in the cost and the profit.

Shiekh says, “You can build an ownership model with decentralized AI. On one side you have this model, you have just trained it and you want to generate revenue from it. Now, let’s say 10 hospitals have trained it in different proportions. Maybe somebody provided 50% improvement, somebody provided 10% data points, somebody provided 20% data points. You can have  built-in ownership of that for data. Machine learning models gives you ownership of that in the percentages that you’ve trained for it.”

With the technology that Fetch.ai is building, our future of AI will be different from what we imagined until now. We are just at the beginning of it. The near future might include centralized AI with decentralized AI propagating alongside it for specific use cases. The convergence of both may be our actual future of AI.

Sheikh says, “It’s the same problem that Google aimed to solve for at the beginning of the Internet revolution. If you think about it, there’s thousands of websites. How do you communicate with them? This is where you need a multi-agent system to evolve into a more autonomous actions framework.”

originally posted on forbes.com by Jun Wu
About Author: Jun Wu is a Hybrid Journalist for Technology, AI, Data Science. She has a background in programming and statistics.