Live data from Hacker News

Machine Learning in the Cloud, with TensorFlow

googleresearch.blogspot.com

11–17 of 17 posts

Re: Machine Learning in the Cloud, with TensorFlow

#12
So if I understand the situation correctly, using this offer, we all feed the AI Google is building, thereby creating another Google monopoly situation (the AI system that gets the most training will be the strongest, thereby attracting even more users/training and becoming even stronger).

Re: Machine Learning in the Cloud, with TensorFlow

#13
post #12

So if I understand the situation correctly, using this offer, we all feed the AI Google is building, thereby creating another Google monopoly situation (the AI system that gets the most training will be the strongest, thereby attracting even more users/training and becoming even stronger).

At the risk of giving a serious answer when your phrasing makes it seem like you're trolling: No. The Cloud ML service is your data and your model and your training results (except as required for google computers to touch it to train it and store it - but the results of that training are yours.)

The same question was asked earlier of the Vision API, with the same answer: https://news.ycombinator.com/item?id=11128444

At a high level, you can think of it as having mostly the same properties as if you rented a bunch of GCE machines (or AWS machines) and ran TensorFlow on them, with your data stored in GCS (or S3 or whatever). The difference is that Cloud ML handles the scaling pain for you -- managing the machines, starting and keeping tasks running, load balancing, etc.

(disclaimer: I'm working on TensorFlow, not Cloud ML.)

Re: Machine Learning in the Cloud, with TensorFlow

#16

Does Google have any academic discounts? I'm using tensorflow for my thesis and wouldn't mind some free cycles :)

Yes. Have your advisor email me (dga at cs . cmu . edu) and I'll send him / her pointers to some of the programs Google has for faculty. Note: cloudml alpha is a limited invite-only thing, so is unlikely to work for you on a short timescale, but nothing stops you from using GCE as a source of cycles on which to run your own install of TF in the meantime.
Post reply on HN