What are the concrete tools to dissect algorithmic black boxes, identify potential biases and gain the confidence of stakeholders? Overview of the methods of explainability to be implemented in your business.
The transparency and explanability of algorithms are as central and crucial issues to establish confidence and ensure a controlled AI deployment within a company. A real headache for the decision-maker who wants to take advantage of the benefits of artificial intelligence, but which fears the return of a stick with a hazardous deployment. Fortunately, the community of artificial intelligence is well aware of these legitimate fears. Thus, methodologies and software tools have emerged to (begin to) provide concrete solutions to this problem.
AI explanatory tools and methodologies
Several approaches make it possible to further illuminate the functioning of AI. First, the fine understanding of the data used to cause models is essential. Practices such as the creation of Datasheets for Datasets or Data Cards offer the possibility of precisely documenting the origin, composition and therefore the potential limits of data games. This ultimately allows you to identify possible biases upstream.
In an attempt to understand the models themselves, the Domaine de l’Explaninable AI (XAI) offers different techniques. Methods like Lime (Interpretable Local Model-Agnis Explanations) or SHAP (Shapley Additive Explanations) promise to obtain explanations on individual predictions, on a case-by-case basis.
To give a concrete example, it is thanks to these methods that it is possible to understand why a system has recommended a specific product rather than another to a customer by identifying the factors that weighed the most in the decision. Other techniques are working to provide a more global view, by looking for which variables mainly influence the general behavior of the model. When the situation allows, it is better to have recourse to intrinsically simpler models, such as decision -making trees.
The major players in the sector now integrate these capacities into their platforms. Google Cloud with Vertex Ai Explainability for example or Microsoft with its responsible dashboard AI on Azure, which is particularly based on the Interpretml library. Open Source initiatives, such as explaining 360 (IBM) or the same interpretml, also provide precious tools to developers.
Finally, rigorous traceability, via detailed logs recording queries, entry data and decisions taken, is a basic methodology to be able to analyze the behavior of a system.
Vigilance points
However, adopting these tools is unfortunately not enough. In an ideal world, this approach should be integrated from the design phase of AI projects. As well as other actions to supervise practices such as the implementation of internal governance and a clear ethical charter.
The explanability remains very complicated to obtain for the most sophisticated models. It will no doubt have to accept a compromise between the pure performance of a model and its degree of interpretability. And to add a layer of complexity to an already simple operation, the explanation delivered by an XAI tool does not always equal to a deep understanding … and can therefore require a truly expert human interpretation not to lead to erroneous conclusions.
These tools constitute precious aid, but cannot replace human judgment and supervision. It should never be forgotten that the final responsibility for the decisions of AI systems will always be the responsibility of the company. Such skills require either logically expensive profiles, or go through external providers. In all cases, the implementation of these processes requires time and very specific skills.
