Microsoft To Launch Machine Learning Service

angry tapir (1463043) writes Microsoft will soon offer a service aimed at making machine-learning technology more widely usable. "We want to bring machine learning to many more people," Eron Kelly, Microsoft corporate vice president and director SQL Server marketing, said of Microsoft Azure Machine Learning, due to be launched in beta form in July. "The line of business owners and the marketing teams really want to use data to get ahead, but data volumes are getting so large that it is difficult for businesses to sift through it all," Kelly said. The service will have "...an interface called the Machine Learning Studio. The palette includes visual icons for some of the most commonly used machine-learning algorithms, allowing the user to drag and drop them into a visually depicted workflow." Algorithms themselves are implemented in R, which the user of the service can use directly as well.

Comments

You want machine learning? You want IBM's Watson.

[mmell]

That machine wasn't merely a decision engine or a huge database - it was a learning system which was given a few months to learn all it could on a broad variety of subjects by crawling the internet. It wasn't programmed with rules to pair an answer with the correct question; it was a system which "learned" how to associate concepts. It was programmed to permit it to weigh its questions against how well they correlated to the answers - to determine a confidence level - but it wasn't specifically programmed to devise questions associated with answers. That's why the next publicly stated idea for the system was in medical diagnostics. It's another area where the ability to relate multiple seemingly disparate items of information with a non-static data store seemed to be of value.

Giving the right programming, it might even hold a conversation better than a 13 year old Ukrainian boy.

Re:Sky.NET

[sribe]

Why did it have to be Microsoft?

Well, at least it will keep crashing, and maybe humanity will figure out a way to defeat it during one of the reboots.

Give WEKA a try

[AlienSexist]

WEKA is Open Source, has an adequate GUI, many different kinds of algorithms available, and a "knowledge flow" visual designer for you to chain it all together. I've used it in a few personal and professional projects to find things like which variables most strongly influence an outcome, decision trees, derived formulas and expressions that accurately predict outputs from inputs, and various kinds of data visualizations for clustering data samples. Code is in Java so I presume you could embed it within a system to automatically perform analysis and swap algorithms on the fly. Best of all, since this is software under your control, and not a Corporate-offered service, your valuable data never leaves your control.

I think WEKA already did a lot to make these kinds of data analysis accessible as Microsoft is aiming to do. No matter who provides it to you, there is something totally awesome about being able to click a few buttons and get some interesting results to munch on.

Re:Give WEKA a try

[lorinc]

I have only one problem with fancy GUI that allow you to train a predicting model in 2 clicks: how confident can you be in your model, since all the parameters are masked and you have no knowledge about them? I still think it is dangerous to rely on a tool you don't understand and you can't control up to a satisfactory level, especially when it is to be used in prediction - something we expect to be highly reliable in many aspects due to old development of science like balistics.

I've written a ML library myself (also in Java, more lightweight than weka, but with no gui - although it comes with standalone binaries for some basic setups) and I can tell you there is no good default tunning that works well for every kind of situation. ML is seriously a young science that gets rapidly tricky even on very common problems, which is very different than field for which we have very accurate solvers that work most of the time (again balistics is probably a good example, at least because it is taught in school and sets the prototype of what we name science). I fear hidding the youth (and thus the imperfection) is only going to cause damage through misconception and false interpretation.

Re:Give WEKA a try

[LetterRip]

Having used Weka, sklearn, vowpal wabbit, Orange, and lots of others - I'd say that Weka is my least favorite, it is slow and has crappy memory management. I do use it occassionaly for feature ranking and reduction, but rarely for the actual machine learning.

Re:Give WEKA a try

[Jon Peterson]

You don't trust anything you haven't built???

How do you know your HDD firmware isn't corrupting data? Build it yourself??

Re:Give WEKA a try

[radtea]

I have only one problem with fancy GUI that allow you to train a predicting model in 2 clicks: how confident can you be in your model, since all the parameters are masked and you have no knowledge about them

This. Expecting people with no knowledge of statistics, programming, or machine learning algorithms to develop appropriate models and interpret them correctly is unrealistic.

I've worked on applications that attempted to bring machine learning to the desktop for specific tasks, and it failed because naive users were simply not able to bring the required knowledge and nuance to the table. It made experts better (I still use the program myself now and then) but there was a threshold below which it was useless.

Without getting your hands on the data and understanding something about the internals of the problem you can't build learning systems that will be plausibly generalizable. Overtraining (and undervalidation) is still a rampant problem in ML applications, and this kind of technology will just make it worse.

Learning algorithms, especially at the level they are implemented in R, are not black boxes. They need to be understood and applied carefully and appropriately.

Re:Sky.NET

[Cryacin]

(Arnie) Reboot this!

Triple E

[fph il quozientatore]

So, R is being Embraced. I wander what's coming next...

Re:Sky.NET

[Jon Peterson]

Google, Apple, Oracle, IBM, etc. etc.

Actually, Azure is great, and the addition of high level services like this is the right direction. Just spinning up VMs isn't nearly as useful as a service layer.

The algorithms aren't an especially hard part of machine learning, dealing with the data is. Anything that would save me the hassle of trying to fit things in RAM would be great...

Re:Sky.NET

[LifesABeach]

I guess we will see R# in the next version of Visual Studio?