Watch out! Rapid Learning--a powerful image machine learning plugin is out!

[Will Ouyang]

I am pleased to announce that a new plugin with full-featured machine learning plugin is out. :)

Thanks to RapidMiner(formerly YALE) , which is a powerful and intuitive data mining tool implemented in Java and available under GPL (GNU General Public License).

The plugin, Rapid Learning, brings the power of RapidMiner to Icy, enables infinite probability, you can do all kinds things such as data loading and transformation (Extract, transform, load, a.k.a. ETL), data preprocessing and visualisation, modelling, evaluation, and deployment.

In Rapid Learning plugin, dimensions of (Z,T and C) are treated as features, packaged as an ExampleSet, the ExampleSet then can be used for training and predicting in RapidMiner process. 

Currently, you can do supervised learning and unsupervised learning to sequence. In supervised learning, you need to define some labels in Mask Editor and draw then on your sequence, then you can train your data and do predicting to other data. You also can do unsupervised learning like clustering, you will use Mask Editor to define your training set and then you can train the model and predicting other sequence.

Of course, you should download RapidMiner first and put the "lib" folder into "plugins" of Icy. Maybe I should make another plugin to include RapidMiner, but the package is too big to a icy library and I should find some way to enables the user use it as a standalone software.

You may also need the example process file needed by the plugin in the attachments. You can find some more information in the plugin page, about how to use them and how to generate it by yourself.

Since it's a preview edition, it works under my tests, there still many work to do,so suggestions are highly appreciated.

Welcome to have a try.

[Will Ouyang]

Hi, 

I just record two videos to demonstrate the usage of the new plugin, you can find here: http://icy.bioimageanalysis.org/plugin/Rapid_Learning .

Best,

Will

[Fabrice de Chaumont]

Hi Will !

What you created is very impressive ! I love love love good documentation ! It is really the difference between a regular and an excellent plugin: people in the end can really use it !

I also put your website link here: http://oeway.github.io/EVA/
I encourage everybody to watch this video: it is just an amazing showcase ! In one single film, Will show his plugins, his open harware, the way it is connected to micro manager, the Icy bridge features with ImageJ, connexions with other plugins, graphical programming, machine learning... and other stuff !

Congrats !

Fabrice

ps: We can advertise for your plugins with an Icy splash screen (or several)
Just use the plugin splash screen maker with one of your image and send it to me !

[Will Ouyang]

Hi, Fabrice

Really happy to see you love it, thank you.

Actually, since no one replied this thread, I am a little confused. I think machine learning and big data related methods will be more popular in bioimaging and other field since some task is really complicated to solve with conventional methods. 

Best,

Will Ouyang

[Stephane]

Hi Will,

In my case i was busy these last days but i am also very impressed by what you achieved and your video is really amazing !
I think that Machine Learning is not that much used "yet" in Bio Imaging, at least it really depends from the context / topic but i believe it will be more and more used with mature tools :) About the plugin itself, it looks like really powerful ! But i think Icy users expect "all in one" tools and the fact you need to install the RapidMiner software aside can be a real "brake" (we already experience the problem with Micro Manager). And this is even more true when the software in question is not free and/or need some registration processes. Of course that is not your fault ;)  but that probably explain why you did not get more feedback about it yet. On our side, we will try to promote it as it really deserves it !

- Stephane

[Will Ouyang]

Hi Stephane,

Thanks.

I have also realized that "brake", I tried to pack the libraries of RapidMiner into an icy plugin, but the file size reached about 50MB, I don't know whether it is allowed. 

The thing I should clarify is, RapidMiner have a community version can be downloaded from sourceforge,it is free and don't need any registration process. 

The newest open-source version is RapidMiner 5, as it official website said:

Whenever we release a new major version, all previous versions become open source. So that means that RapidMiner 5 stays open source (free) now that we have released version 6.  And, when 7 is released, 6 becomes free and so on.

RapidMiner is distributed under AGPL, so it is possible to pack the libs to make a lib plugin, if we don't mind its size.

Actually, I have investigated about which machine learning package is suitable for Icy, there are many choice, including Orange, R, Statistica and WEKA and so on. There also some article about the comparison, like this one. Eventually, I choose RapidMiner, because it is powerful and polish as Icy, the protocol like programming style, and even WEKA operators can be used by RapidMiner.

I think it's right choice for Icy.

Another reason for choosing RapidMiner is the RapidAnalytics project, this project enables us to establish a cloud a machine learning server and run machine learning task in the cloud. This will be useful if we want to build a knowledge base in the cloud in the future. Online machine learning will be implemented in RapidMiner, users can feed their data to train the online classifier, and finally it may be able to complete task like counting cells,and accurately like a human.

For further development of the plugin, I think I should integrate some common purpose classifiers and clusters which are ready to use when opening Rapid Learning plugin. Classifiers act as CellProfiler Analyst will be able to build, even better result can be achieved. It is also possible to make some already trained classifier to do some common task, and maybe it can be packaged as a plugin by "Python Plugin Packager" (thanks to Timothée's job),  the only problem is we should find some mechanism which handle all kinds of images feed by user, the motivation for this is that sometimes the training process is really tedious and  even not possible for a common user.

Cheers,

Will Ouyang