Understandable Machine-Learning Systems to Benefit Everyday Routines

Research Areas: 

Systems and artifacts that use machine learning are going to become ever more important in our every-day life. They come with the disadvantage of being a black box that does not convey their reasons and can not formulate their behaviors in a human-understandable way. Besides making systems behave in an intuitive, that is predictable, way, another approach is to give them the capability to elaborate their input/output relations in a symbolic instead of the poorly understandable statistical way in which most modern machine learning techniques represent their knowledge.


Methods and Research Questions: 

Besides giving systems the possibility to explain themselves and making themselves predictable through so called rule extraction, there should also be the option to change the learned mapping using the same symbolic interface. This would have several advantages, the foremost of which would be that the world knowledge of the users about their own daily routines and personal life can be taught instead of learned. We also expect such a system to work on a psychological level, giving users back a feeling of control that they otherwise lose through the black-box character of subsymbolic systems.

Lastly, users might be constantly think about the effect of training data they give to a adaptive system, being afraid of spoiling it with skewed examples which they know not to be representative. Being able to look at what a system learns and correcting it if needed would probably much improve the user experience as well as the learning success.

We are currently testing these hypotheses with an exemplary sensor fusion system for context prediction. In Figure 2 you can see how it predicted the availability of a user, using an instant messenger as both input and output device. Therefore it is tied in tightly with the already exhibited behavior of many computer users, giving us natural data in a relatively simple environment and a functionality that is immediately useful for users. It has a very minimalistic interface in order not to distract the user unnecessarily (cf. Figure 2). The rule extraction is done using the G-REX framework by König et al. [1] and as of now presented in a simple tree structure shown in Figure 3.