louisdorard

I created the Machine Learning Canvas to make it easier to ask the right questions at the beginning of an ML project, and to save people from wasting time and money due to a poor design of their ML system. I’m now releasing the first draft of a book that contains everything there is to know about this framework, in a 1-hour read.

You've probably heard about Amazon Machine Learning or Google Prediction API or BigML, but as a developer, how do you choose one of these APIs when you want to integrate ML in real-world apps?

Most of them can. I have a doubt for Google Prediction, need to check... Also with BigML you get a decision tree model which allows to "explain" predictions with a list of decisions based on the values of the fields (see #3 on https://bigml.com/features).

Have you guys been able to try both and compare them?

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 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?

Sounds very exciting! This space is getting a bit crowded though (see http://www.quora.com/Who-are-the-main-competitors-to-the-Goo...). How would you differentiate Datapal from competitors such as BigML, Predictobot, etc.?

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.

> "this is more in line of those graphical tools that claims you can create programs without having to learn programming"

There's a comparison to be made with graphical tools that let you create websites without knowing html/css/javascript, like squarespace.com for instance. It's enough to cover people's needs in 80% of the situations. See this article by Scott Brave for more: http://gigaom.com/2012/12/22/we-dont-need-more-data-scientis... .

Have you tried using BigML? (which to me is extremely similar to what Azure ML is pitching)

Prediction.io needs a server to run on, and it focuses on recommendation problems, so it doesn't seem to do classification and regression for instance.

Reddit entrepreneurship?

Monte Carlo Tree Search uses Machine Learning indeed! The balance between exploration and exploitation of the game tree is achieved using "bandit" algorithms (a type of Reinforcement Learning algorithm). I recommend reading (Kocsis & Szepesvari, 2006): http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.102....

cheers speedraf :)

I'm still wondering how many potential users they're losing by forcing registration and having a "software" type of website...

I might have had the wrong idea. I am very interested in examples that point to the contrary!

You're welcome qwerta!

I want to stress that I am only talking about my own experience and this isn't a study on SV. The comparisons I make are also based on the various tech meetups and events I have been to (unrelated to VA-Live).

Thanks sn0v!

Hey James, thanks for your feedback. Which part of the US do you think we should have gone to instead? Are there any conferences and professional bodies you could recommend us to check out?

Just changed the title based on everyone's feedback. Thanks for reading anyways!

Just updated the title based on everyone's feedback — thanks for your suggestions. Will update the blog itself soon!

Thanks for backing me up bennyg :)

Thanks Iwan. Do you think I should change that title after the link has been posted?

My intention was to have people really question whether it is the right time for them to go. Anyone else think I should change that title? If so, I'm happy to experiment...

Is it ok to change a title after the link has been posted?

I have to admit that the title is provocative. I do think that going to SV can be a waste of time/money in certain situations, but hopefully those tips can help :)

Oops, sorry if the title was misleading! I was trying to point out that there are things that are not related to the algorithms being used that can make learning fail, namely the way that data is collected and noise in the observations.

Here's what it's like to be an authority ;)

This is exactly how I see these new ebooks, not so much as actual competitors to the oreilly kind of books (because they are more concise), but as longer form thinking than blogs. There's a lot of enthusiasm on self-publishing books, and maybe it will be short-lived, but I hope that in the long term we'll have more and more experts in our community who will share their knowledge on niche topics in the long, self-contained and polished form that books are!

That's a very good point. These 1000 words often get replaced or deleted. I guess that in many creative activities it's painful to edit down, but it usually pays off. Sometimes you don't just delete what you wrote but you rephrase it. I find it much easier to formulate a complex idea when I already have a first version in front of me (even if it's terrible), than to start from nothing.