RE: Testing Machine Learning tools for optimizing the Steem experience
Great, I was just talking about it to Alex while watching a great presentation:
We need an n-gram model to improve the YouTube output to account for unknown words like names like for example BitShares and DEX. An appropriate language model is the best sollution to the current transcription problem because of the different and highly specific terms we use in our crypto space.
KenLM: kheafield.com/code/kenlm
What you do is create a LM for countering weird stuff. They call it the Tchaikovsky problem.
Training time on an TitanX ~= 30 days We can feed it all the Mumble talks from the past years, it needs about 10.000 hours of training. If we feed it a good language model by leveraging steem, bitsharestalk, bitcointalk then we'll have a better engine than YouTube. Can use gridgoin network for this and have it in less than a week.
You should contact admin of Citizen Science Grid. They are experimenting with neural learning on boinc. You will soon run into a problem: If it takes 30 says on a TitanX (nvidia), are you able to split the task to hundredths of computers?
Thanks, I have no idea, well there are also the #gridcoin people I could ask...
that could possibly be a solution, if you really wanted to give it a try ... https://boinc.berkeley.edu/trac/wiki/PythonApps
Thanks Alex but Boinc is a volunteer project, so you mean to add a new project? I would have thought peope already had a speech-to-text machine learning task running on Boinc but I don't see it in the list of projects...
I was just saying it should be possible. I believe though (if I recall correctly) new projects would have to be proposed to the community, and you have to get people interested in dedicating time to your project.
However, this may be a better question to direct to @tomasbrod, since he originally suggested boinc/gridcoin at the top of this thread!