As machine learning requires more and more computing power, people turn to services such as AWS, Google Cloud, or similar to train tedious machine learning algorithms.
While those services are nice, they are also pricey and you have to resolve that your data will be detained by big companies...
LearningChain creates a network of computers (the graph) running its client that allows individuals as well as universities and companies to share their computing power in a fair way, in order to be able to make others do computations for you.
Although the service is completely free, you need to make computations for someone else first to acquire credits that will allow you to submit your own task to the graph.
When a task is submitted, every client can grab it, but only the best answer will receive the credits for it.
We chose the blockchain technology to have a complete history of all that happen over the network, as well as to manage the credit system. A python client takes care of handling data from the blockchain : it receives a new job from the graph, runs it, then announces its result to the network, and if it is the best one, gets paid and returns the computed model back to the contractor.
In order to avoid harmful code to be executed on your computer, the learning is done in isolation in a docker container that can not access the internet nor any other resource that may exist in your computer and that was not included in the contract.
So here is exactly what happens when you want to use our service.
- You download our client and run it. It will generate an ECC key pair that will uniquely identify yourself on the graph
- You configure the client to run only jobs that you have interests in : validity date far enough, bounty high enough, and let it run until you've acquired enough credits.
- You create a docker container with everything your model needs for its training. You host it on the container host of your choice.
- You create the job with the bounty you are ready to pay for its completion. The higher the bounty, the more the graph is interested in your job, and the faster the result arrive !
- Your job is propagated on the graph and the bounty taken out of your balance.
- As soon as a node founds a solution, it propagates it to the network, announcing the computed loss.
- Anyone can still try to beat that announced loss until the expiration date of your job.
- When your job expires, the node that gave the best loss wins the bounty. The news is propagated on the graph. If more than one node have the exact same loss, the fastest one wins.
- The winner propagates the computed model on the graph.
- Every node can check whether the winner lied on the announced loss by trying the proposed model. If enough node consider the winner to be a cheater, it is banned from the graph and its credits are lost. If this is the case, your bounty is then returned to you.
Building a blockchain from scratch is hard, really hard, and we had to solve a lot of issues towards a blockchain systems that is resilient to malfunctioning nodes as well as malicious ones. It took a long time designing a protocol we can trust over the graph before we reached a consensus.
After designing, implementation was interminable. It took us all night to code every aspect of our proof-of-concept, but we are proud of the result.
Well most basic one is : we've done it ! Really, the coding power of 4 motivated guys during more than 20 hours was the very minimum we could have performed such a task with.
We also are thrilled by the whole protocol we came with. Of course we have no proof it is truly safe, but we believe that all the most basic attacks you can think of will be unexploitable on LearningChain's network.
This project forced us to tackle many problems we were unaware of when designing a decentralized distributed system. It also taught us that sometimes, the solution is so simple that the only way to see it is to take a step back and rethink everything.
Well, this is still a proof of concept. We would need to actually develop a node communication protocol for the blockchain and a discovery service, and we would also need a thorough test of the network to be sure everything is as secure as we can make it.
We decided to tackle the challenge issued by AdNovum