So many ML libraries and algorithms…
In the Machine Learning field there is a great wealth of well-studied approaches to various learning problems and users often experience the agony of choice when having to decide for one specific algorithm for a given problem. Algorithms may differ significantly in their runtime behavior, their memory and data requirements and ultimately their final quality. The precise choice of the most appropriate model may vary from situation to situation: What seems like a good setup when data is abundant and training time can run for many hours will fail to satisfy your needs if only few samples are available or if a working model is needed within moments.
Offering a range of algorithms to choose from therefore is a good idea for a machine learning platform. In order to allow users to explore different options and compare their respective benefits, a welcome way to offer that is to ensure that the various options accept their training data in the same format and deliver their results in the same format, too such that results can easily be compared. The python library scikit-learn for instance features a rich selection of learning algorithms underneath a uniform programming interface, making experiments with approaches such as decision trees, naïve bayes, random forests and many others relatively simple. However, users still have to be somewhat fluent in writing python code and ensuring that data from the various experiments ends up in a similar format for subsequent comparison.
A ML models experimentation framework
The Kairntech Sherpa platform similarly gives users access to various powerful machine learning packages, “hidden” underneath a simple and easy to use user interface but in addition eliminates the need to engage in coding. The Kairntech Sherpa platform currently gives the users access to a simple and fast CRF (Conditional Random Field) approach “CRFSuite” as well as to the Spacy package and Delft, a deep learning powered package that is available in the public domain and that has been designed by Kairntech Chief ML expert Patrice Lopez. By adding more and more such packages, we enable users to experiment with and benefit from the package that is most appropriate for their respective task without having to worry about the necessity to dive into writing code or the different conventions for installing, employing and training them.
We are currently in the process of adding Flair too, another powerful deep learning library. We are happy to present these new features in the Kairntech Sherpa platform on the well-known Conll 2003 corpus for named entity extraction. The Kairntech Sherpa platform supports running such comparison experiments, allowing to directly home in on the most promising setting.
The Kairntech Sherpa platform keeps track of the experiments on a given data set that users perform using different algorithms or different settings and offers detailed reports about the respective outcomes. And here it is not without some satisfaction that we notice that the Delft package mentioned above turns out to be pretty competitive on the Conll 2003 benchmark data – see the report above.
Kairntech maintains instances of the Kairntech Sherpa platform online for test, demos and experiments. Feel free to contact us at info@kairntech.com to discuss your first steps with the Kairntech Sherpa platform.