Like most other exponential growth patterns, there is suddenly HUGE number of people now wanting to use ML but don't have necessary background. Pretty much every other enterprisy developer who had been happily spending their lives pushing data from RDBMS to/from UI now wants to try out ML for something or other. Unfortunately current state of ML framework is a minefield. If you don't know your precision from recall or confusion table really confuses you or have no clue when you have been merely overfitting all along, you are screwed.
I can see huge market if you can bottle down complexities of ML in a nice easy to use package. A very easy to use ML service or tool would allows you to import your data with few clicks, apply normalization such as stemming or edit distance by pick-n-choose without writing any code, has nice generic featurization library, can run tons of algorithms with large number of parameter sweeps, does automatic feature selection, takes care of maintaining test and validation sets etc. This kind of thingy would be super popular. My guess is that it has to run in cloud because it eliminates all the setup and updates plus doing what I described even on moderate size data sets usually takes hours on single machine. Now throw in plug-n-pray deep learning algos which even few ML experts are familiar with and can require significant GPU based infrastructure.
Can this be reality? In my opinion, this is more in line of those graphical tools that claims you can create programs without having to learn programming. It's impressive when you do demo but don't last long in battle grounds. Ultimately, if you want to write program, you will need to get your hands and cloths dirty with oil stinks. If you want to do machine learning, you will need to start from stats and probability.
When I said I'm on the other side, I met I'm one of those crazy machine learning guys[1] who doesn't see a business model in the full blown package.
You're right that this would be hard, but I'm wondering if the cost benefit analysis is there for it. I looked at this approach, and I don't believe it is relative to the costs incurred for something like this. I'd be willing to give it a shot at some point maybe, but neural nets behind the scenes for me involves more fun data behind fire walls ;).
> "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... .
I can see huge market if you can bottle down complexities of ML in a nice easy to use package. A very easy to use ML service or tool would allows you to import your data with few clicks, apply normalization such as stemming or edit distance by pick-n-choose without writing any code, has nice generic featurization library, can run tons of algorithms with large number of parameter sweeps, does automatic feature selection, takes care of maintaining test and validation sets etc. This kind of thingy would be super popular. My guess is that it has to run in cloud because it eliminates all the setup and updates plus doing what I described even on moderate size data sets usually takes hours on single machine. Now throw in plug-n-pray deep learning algos which even few ML experts are familiar with and can require significant GPU based infrastructure.
Can this be reality? In my opinion, this is more in line of those graphical tools that claims you can create programs without having to learn programming. It's impressive when you do demo but don't last long in battle grounds. Ultimately, if you want to write program, you will need to get your hands and cloths dirty with oil stinks. If you want to do machine learning, you will need to start from stats and probability.