The human factor

Additional difficulties in obtaining benefits from the use of AI arise as a result of employees' distrust of recommendations coming out of the "black box". People tend to make mistakes and be confident in their own rightness, so they are more likely to believe their experience, knowledge and intuition than an AI that is not able to explain its own decisions. As a result, the technology can be ignored by employees. Therefore, it is important to regularly test algorithms, which will not only identify and then eliminate their shortcomings, but also demonstrate their effectiveness to employees.

Speaking of testing, it is useful to ask how representative its results are. As a rule, testing looks like this: conditions are created that are close to real, in which the algorithm will have to work; the algorithm starts, gives forecasts or recommendations; finally, performance indicators (metrics) are calculated. The quality of the algorithm is evaluated by the values of performance indicators. What's the catch? Firstly, the conditions in which testing takes place may be far from reality, which will not give an objective assessment. Secondly, the performance indicator is chosen either by a mathematical developer who may not fully understand the business process and give priority to mathematical metrics; or by a business customer who, on the contrary, may give excessive weight to some aspects of the algorithm, but not taking into account others. The optimal indicator should be sought at the intersection of two worlds, but the proportion of subjectivity will still remain. Thus, the choice of testing methodology is very important, and it falls on the shoulders of the developer and the customer – living people.

Another of the myths about AI says that machine learning algorithms are trained independently – only submit data. This is not so: some elements of their design, again, fall on the shoulders of the developer, on his expert opinion. A classic example of such elements are the meta-parameters of machine learning models. Of course, there are methods for their automatic selection, but even in this case, the choice of a specific method and its parameters remains with the developer.

#yurysirota #AI #artificialintelligence #cdo #chiefdataofficer #chiefanalitycsofficer #bigdata #dataanalytics #datascience