Transparency is one of the most frequent catch words when discussing ethics in relation to AI and machine learning. The background is that the responses delivered by learning machines have been argued to be a delivery out of a 'black box'. But is there any work going on with the focus on making the process more transparent (I assume something like making AI systems not only produce answers or recommendations, but also explaining and supplying support for the solutions presented)?
Of course there are people doing so, here you have an agregattor of python libraries https://github.com/tensorflow/lucid#notebooks
This Collab notebook ilustrates what you are looking for https://colab.research.google.com/github/tensorflow/lucid/blob/master/notebooks/building-blocks/SemanticDictionary.ipynb
I have to say that all these techniques still need a lot of improvement
Hi Gert Helgesson ,
while I think this is a very interesting topic, I have doubts about a general success, based at least on two reasons:
Part of the AI approach are expert systems. However, "experts" as defined here are not necessarily humans who have a complete understanding of a complex electronic circuit, for example, starting at the properties of all components, the laws of electromagnetics, the detailed circuit, and so on, but people who know from experience rules like: If LED3 is blinking replace board X. So, if you could ask an expert system which has learned these rules from a human expert why it suggests to replace board X, it could answer only: "Because rule 73 says so in case of a blinking LED3." So, the transparency reveals just a lot of little black boxes inside the enclosing, formerly black box.
Secondly, I think a general rule is: If you perfectly know the algorithm a process is based on, just write a program according to this algorithm. If not, try to implement the process by machine learning. If machine learning is successful, obviously a solution has been found but neither the neural network (NN), for example, nor its architect know exactly why. Graphical representations can be certainly very helpful on a scientific level, but don't provide answers to the uninitiated user.
What is more, today in some cases NNs are applied in a "magical" way instead of being based on rational ground; take, for example, predicting the stock market: The correct answer, based on real world facts, to the question: "Why will the shares of company Y go rapidly downhill?" were: "Because company Z has made an invention that renders the product of company Y useless." but the NN knows exclusively spectra of and correlations between a lot of abstract data.
- Badges
- Science topic