No account yet ?

The famous "black box" of contemporary predictive artificial intelligence, that is to say the lack of transparency in the decision-making process of predictive AI algorithms, poses a real problem for the massive adoption of this technology ( Predictive AI).

Often, deep learning algorithms are singled out.

Appreciated for their ability to ingest large volumes of data and their efficiency in solving complex problems, they are often seen as black boxes. Without supervision during learning, decision-making processes seem complex to trace. More and more research teams are looking to put in place methods to more easily control these neural networks without success.

The different types of bias to be aware of :

Sampling bias. One of the most common mistakes in data collection results from a lack of representativeness. Some elements may be oversampled from reality. Take the example of a company that tries to predict machine failures. If it mainly collects information on errors, the algorithm will not be able to identify sufficiently precisely the normal operation of the equipment.

Measurement bias. Measurement bias is the result of the failure to accurately measure or record the data that has been selected. For example, if you use salary as a measure, there may be differences in salaries (bonuses, benefits…), or regional differences in the data. Other measurement biases can result from the use of incorrect data normalization or calculation errors.

Exclusion bias. Like sampling bias, exclusion bias arises from data that is inappropriately removed from the data source. When you have petabytes of data, it's tempting to select a small sample to use for training, but doing so may inadvertently exclude some data, resulting in a skewed dataset. Exclusion bias can also occur due to the removal of duplicates in the data when the data items are truly distinct.

Registration bias. Sometimes the act of recording the data itself can be biased. When recording data, the researcher may only record some data and ignore others. You can design a machine learning algorithm based on data from IoT sensors, but if the recording is not continuous, some data may be missing. Or there is another systemic problem in the way the data has been observed or recorded. In some cases, the data itself can even become biased by the act of observing or recording that data, which could trigger changes in behavior. Other problems can arise during this collection phase. They can cause behavioral changes of the algorithm.

Bias related to prejudices. In some cases, the input data is tainted with human biases that will favor different elements based on their ideology. When using historical data to train models, especially in areas where prejudices were previously widespread, care should be taken that new models do not take them into account.

Confirmation bias. Confirmation bias is the desire to select only information that supports or confirms something you already know, rather than data that might suggest something that goes against preconceptions.

Trend. This form of bias generally appears when data scientists spot a trend in a data set. As the volume of information relating to this trend increases, there is a risk of over-representing this phenomenon, which may ultimately be short-lived.

And many other ways too !

There is a plethora of articles on the subject on the web, you just have to search for the expression "les biais de l'ia" in French and / or "AI bias" in English without the quotes on the search engines.