Much of the discussion about blockchain technology to date has either lauded its transformational promises or snickered at the marketing hype undeterred by the lack of actual breakthrough applications. But Algorithmia, a startup curating a marketplace of machine-learning models, just demonstrated how Ethereum can be used to submit and verify those machine-learning models in a recent contest.
Back in February Algorthmia introduced a contest around its new Danku protocol, which was designed to see how well the Ethereum blockchain worked as a vehicle for verifying machine-learning models without Algorithmia’s active involvement. Last week the company revealed that a winner had emerged, who submitted a model predicting how a given area of the United States was liable to vote based on a limited random sample of 2016 election data with the required level of accuracy.
“We’ve been thinking about, how we do acquire machine-learning models faster and in a better way?” said Diego Oppenheimer, Algorithmia founder and CEO (and a GeekWire Cloud Tech Summit advisor). The idea for Danku was to create a smart contract that automatically verified a machine-learning model as compliant with a pre-determined set of criteria for accuracy coded into a smart contract, which required the company to actually run those machine-learning models on the blockchain.
Contestants were given a random sample of 500 county election results labeled only with their latitude and longtitude, and the task was to create a model that forecast the likelihood of any given county landing in the Red America or Blue America camp. The winning model managed to do so with 83 percent accuracy, and you only had to be 50 percent accurate to qualify; “we weren’t trying to beat Nate Silver on this one,” Oppenheimer joked.
The point of the experiment was to see how machine-learning models performed on the Ethereum blockchain, since the model had to be run in its entirety — and therefore verified — in order to be accepted by the blockchain. The winner noted that the constraints of the blockchain required them to build a relatively small neural network to conform with the terms of the contract, and that could be modified in the future depending on how much computing power the parties want to throw at a problem.
Lots of companies are trying to find a use for blockchain technology beyond cryptocurrencies, from distributed supply-chain services to cloud storage. Machine learning is a field as nascent as blockchain, and there might be promise in the notion of a ledger comprised of verified-as-accurate machine-learning models that others could use in their applications.