Drag-and-drop machine learning

drag-and-drop machine learning love and hatePosted on March 27, 2017 by Lili

Article directory [hide] 1. Past Life 2. Love 3. Hate 4. Summarize

Drag-and-drop machine learning is a problem I've been thinking about for a long time. 1. Past Life

Drag-and-drop machine learning is that people are dragging and dropping on the interface to build a machine learning process. Drag-and-drop machine learning systems typically have a rich set of components, including data cleansing, feature selection, training, forecasting, and performance evaluation. In a way similar to building blocks, people will synthesize a learning component into a completed machine learning process.

The prototype of the drag-and-drop machine learning came up very early. Weka is a data mining software developed by the University of Waikato, New Zealand. Weka provides a Weka Explorer graphical interface In addition to the Java API. People can easily load data, observe data characteristics, train, predict and evaluate the effect by mouse operation on the Weka Explorer interface. In addition to Weka freeware, the business software Matlab and SASS also provide a graphical interface. Personally, I think the software is positioning itself as a software, not a toolkit or a system. But the software is really the first machine learning system with a graphical interface.

With the recent years of machine learning becoming an academic, "Everyone can use machine learning" has become a vision of many people. In their imagination, data preparation, training of different algorithms, prediction of different algorithms, and effect evaluation are all encapsulated in the components, so people can use machine learning smoothly by dragging and dropping the components with a few mouse clicks. Adhering to this philosophy, people have developed a number of drag-and-drop machine learning systems. Among the more popular are the big data packages from Microsoft's Azure machine learning Studio and Ali.

In addition to large companies, there are startups that develop drag-and-drop machine learning systems. The following figure is a deep learning platform based on the Aetros Theano. Users can simply drag and drop to complete a basic structure containing convnet,fcnet.

2. Love

Drag-and-drop machine learning uses the threshold of machine learning, from programming to component drag and configuration file authoring. Machine learning is difficult to use to achieve a qualitative decline. But I have always doubted this advantage. Financial companies, foreign trade companies, banks, and even the Internet enterprises and other organizations, non-technical staff really have the use of machine learning needs and knowledge reserves. I doubt that.

Drag-and-drop machine learning can be a great place for engineers to do machine learning tasks, even if they can't achieve "machine learning for Everyone" beginner's mind. An engineer's machine learning task of organizing funds on the interface may have an intuitive understanding of their machine learning tasks: What step to take for their machine learning tasks, where to go if an error occurs, and which tasks are affected by the error step.

For example, we can see from the image above that if the normalization error occurs, the split and successor tasks will be affected. 3. Hate

said the drag-and-drop machine learning good, we say that the drag-and-drop machine learning bad.

In drag-and-drop machine learning, component plus configuration replaces programming as a way for people to use machine learning. But the components are configured, but not as programmed, to completely deal with the complexity of machine learning. In addition to the understanding of machine learning algorithms, machine learning uses the most complex parts in two parts: Tuning features and tuning parameters. Features include: which features to use, which features to discard, and which feature preprocessing methods (such as scaling) are used. The tuning parameters are related to the specific algorithm, such as logistic regression with the learning rate and the regular factor of two parameters.

Components can be configured in a way that configures a set of feature engineering schemes and a set of parameters, but it is difficult to quickly verify which set of feature engineering scenarios and parameters work best. In the way of programming, we can iterate through the different feature engineering schemes and parameters, and get the corresponding effect index. However, in the way that the components of the drag-and-drop machine learning are configured, we can only remember the different feature engineering schemes and parameters in the document, choose one of them to the drag-and-drop machine learning system, run a few hours to get the evaluation indicator, record the evaluation indicator to the document, and then select the next group. Repeat the above steps until you have traversed all the feature engineering scenarios and parameters. The goal of our programmers is to use code to string together different tasks to automate. But now the drag-and-drop machine learning has rudely fragmented this automation chain.

Then the automation is implemented directly with component plus configuration. If this is to be done, drag-and-drop machine learning provides conditional judgment components and loop components, as well as defining a set of criteria for feature engineering schemes and parameter changes. It seems to be developing a new programming language. This goes back. 4. Summary

Drag-and-drop machine learning "Everyone can use machine learning" Beginner's mind I am not optimistic. For engineers, drag-and-drop machine learning is also full of love and hate.