Watch all the Transform 2020 sessions on-demand here.
Algorithmia announced an AI competition today that could usher in a new era of machine learning model development using the Ethereum blockchain.
The company is offering 3 Ethereum tokens (worth more than $2,500 as of this writing) to the team that can provide the best algorithm for determining voter preferences in the last presidential election, broken down by latitude and longitude. That contest, which has been coded into a smart contract, will last for 60 days. The first team to submit the most accurate model will be the winner.
This competition is really a proof-of-concept test of a system that could allow anyone to create their own smart contract and solicit a custom machine learning model to solve a particular problem. That could help organizations that want to apply machine learning to a particular problem but don’t have the resources to hire a data scientist. Algorithmia’s method doesn’t require participants to trust one another (since all of the components are controlled by the contract), and it automates reward payment.
The DanKu contracts Algorithmia invented allow a developer to outline a particular problem they want solved through a machine learning model, like determining whether or not a transaction is fraudulent, or if a photo contains a chicken.
June 5th: The AI Audit in NYC
Join us next week in NYC to engage with top executive leaders, delving into strategies for auditing AI models to ensure fairness, optimal performance, and ethical compliance across diverse organizations. Secure your attendance for this exclusive invite-only event.
Each task has a minimum accuracy requirement, a reward, a date when competition ends, and a dataset associated with the problem. Creators then use the smart contract to establish randomized pools of training data, which are publicly available for data scientists to apply to model creation, and test data, which is kept secret and used only when checking how well models perform.
That split is important because it helps ensure developers don’t overfit their models to the training dataset. (Overfitting models have been optimized to predict the results of a training dataset, at the risk of reduced accuracy when tested with data they haven’t been trained on.)
All of the computation to test a model is run by the machines powering the Ethereum blockchain. That gives both model and task creators clear verification of results, without requiring either to trust the other. Once a winner is chosen, the contract’s creator automatically gets the winning trained machine learning model, and the developer receives their payment.
DanKu also represents a momentous occasion for Ethereum. Algorithmia cofounder and CEO Diego Oppenheimer said in an interview with VentureBeat that his company believes their development of the DanKu system (named after two Algorithmia engineers who created it) was the first time neural network inference had been run on top of Ethereum.
Oppenheimer isn’t trying to overthrow Kaggle, the Google-owned platform for data science competitions that has hosted such big-ticket events as a million dollar competition to build the best algorithm for Zillow’s Zestimate home price estimation. Instead, the DanKu contracts are designed to make it practical for organizations to farm out development of particular machine learning capabilities to the crowd.
In the future, Oppenheimer envisions a situation in which applications could use this mechanism to autonomously request models from other humans and applications in order to improve their own capabilities.
“If a machine knew how to define a contract automatically, and put it on some system that other systems could solve, there is a potential for machine learning systems solving problems autonomously at a certain level,” he said.
Oppenheimer said that the announcement isn’t tied to an initial coin offering (ICO), nor does Algorithmia plan to create a new cryptocurrency as a result. The company benefits from this arrangement because it operates a marketplace for trained machine learning models, and it also offers companies software to run trained models at scale.
But before the future is upon us, there’s still plenty to figure out.
“There’s a ton of work to do, there’s a ton of things that still need to be moved forward, but that’s the exciting part,” Oppenheimer said. “Okay, so this is possible, and it was done once with real code, now let’s see what the world brings.”
