I think there is a big craze with regard to ML. Most people just draw a black box call it ML/Brain/"hire data scientist" without realizing that for most nontrivial problems its not going to work like magic. Some of the things that I see people underestimating is; a) how hard it is to make the magic black box b) amount of data you need c) how clean the data needs to be d) at times, you need lots of human annotated data -- cost and time to collect it e) how much it's a art than a science. Moreover, people don't have a good understating of the technical challenges and cost when you need to scale to n (features) * m (customers) * p (products). Maybe Azure solves some of these challenges but quickly looking at the pricing -- it seems like its a bit too costly.
I agree that making the magic black box is very hard. Automatic model selection, at scale, requires a lot of processing power.
My take is that Azure "just" makes it easier/quicker to run ML experiments, and to deploy models. It's not entirely black box since you have to pick an algorithm and parameters. I expect that once you've run your experiments and found what works best, you should be able to get similar results with an open source implementation of your chosen algorithm. But then you'd still have to deploy your model somewhere — maybe using a platform like yhathq.com which makes things more transparent?
completely agree,we are facing the similar problem. But Azure provide a minimum viable lab before you take next step towards ML in your company or product stack.
The main goal of this project is to make it easy for people without knowledge in predictive analytics to use their stored data in order to make predictions.
__It is still a very early prototype__, therefore the "predictive power" is not great yet, and all kind of bugs are expected.
I have a lot of ideas on how to improve the service and your feedback would be really appreciated in order to prioritize the next steps!
Hey, thanks for your interest! Well, the market is very big and I think so far all the competitors use different approaches. DataPal will soon include a number of features that will make it more different. Stay tuned! ;)
I've thought about doing this myself, but it's a lot of work. I applaud your effort. I have a lot of experience with internet marketing, perhaps I can help out.
Hey Dan, I've written a book about services that abstract away the complexities of ML: http://louisdorard.com/machine-learning-book I can use all the help to educate people to ML and what they could be building :) Let me know what you think!
I think saw a couple of differentiators from Google's prediction engine. Would someone happen to have a more comprehensive view of how this is different from Google Prediction API?
The main difference between Google Prediction and Azure ML is that the former doesn't require any knowledge of machine learning algorithms, whereas the latter does. Google Prediction automatically selects the best algorithm based on the data you uploaded. In Azure ML, you have to choose an algorithm (and its parameters) yourself.
Other differences are that Azure also has a data transformation component, a built-in text analysis tool, it can perform clustering tasks, and it makes it easier to expose your trained models as APIs.
If any YC startups are interested in checking out Azure, note that Microsoft grants $60k in usage credits (as well as architecture and engineering support). Shoot me a mail at felix.rieseberg@microsoft.com and I'll help you out!
Hey Felix, any chance that that offer is open to researchers as well? One of the groups I'm involved with at UW needs a decent amount of compute power and would love to move some stuff off our homegrown cluster of spare hardware.
There's also the Azure Machine learning university program with videos etc., but looks like it's for partners only, perhaps will be available to all once it's out of preview.
