Hacker News Reader

Question: are there any efforts to communally create / crowd source training material for neural networks?

I’m thinking language like this, but also labelled imagery for, for example, face detection, which works better on white people.

Has anyone attempted to create a way for people to create and donate labelled data to a dataset?

View on HN

This project is more on the academic side -- disclaimer, I was involved in it, but it's led by Justin Harris at Microsoft Research: "Sharing Updatable Models on Blockchain" https://github.com/microsoft/0xDeCA10B

The idea is that the smart contract is a learning algorithm, and people can donate data to a public repository stored on a blockchain. The learned model is publicly available for everyone. People can also receive incentives for donating data in some implementations.

Common Voice ( https://commonvoice.mozilla.org/ ) is one such project. I've urged the people I know who have less-common accents to contribute, but I'm not sure they have.

Good, high quality, wide coverage, labeled datasets are expensive to assemble. Most companies don't want to give them away. You can find a number from academia, though.

Reply on news.ycombinator.com