ROBUST, INTELLIGIBLE & EXPLAINABLE AI

A fuzzy “IF…THEN” rule is a local, non-linear model linking nuanced variables. Mathematically, it defines a non-linear multidimensional function linking some input variables to the output variable.

An xtractis model is defined by a set of fuzzy rules covering the operating space. Any situation/case in this space leads to the simultaneous and gradual triggering of some rules, then to the interpolation of their decisions: the local rules interact with each other and cooperate to calculate the most appropriate final decision.

The higher the number of fuzzy rules of a model and the more input variables it uses, the more it will describe the behavior of a complex process in an accurate way. The prowess of xtractis is to find the actual level of complexity of the studied process: Its ultimate objective is always to discover the most robust and the most compact model, i.e. the most efficient and the most explainable one (see questions 9 & 10).

The xtractis AI was originally invented to model benevolent human behavior: sensory perception of products, consumer liking and subjective evaluation. It was then successfully used to detect malicious human behavior (fraud, security, cybersecurity).

Actually, its universal approach makes it possible to model any complex process or phenomenon including natural phenomena in science… as long as a dataset is available.

Benefitting from the advantages of the RF-N Theory, the xtractis AI is particularly efficient in a complex environment in which the three aspects of fuzziness coexist (imprecision, uncertainty and subjectivity). And this is usually the case of real-world processes that we are trying to model. xtractis robots deploy an infinite family of inductive learning strategies to provide robust models, i.e. models that guarantee the reliability of their predictions in situations that have not been part of the learning phase. Therefore, the xtractis AI avoids overfitting or overlearning.

In addition, these robots automatically and collectively improve their reasoning strategies to continuously improve the robustness of the discovered models, even if the database does not change: they learn to learn better.

The xtractis AI has won many benchmarks against other open source techniques, whether statistical or AI-based: Polynomial Regression, Logistic Regression, PLS, CART Decision Tree, Random Forest, Boosted Trees, Kernel Support Vector Machine, Deep Learning / Neural Networks… Of course, INTELLITECH’s team has an expert knowledge of all these techniques to be able to highlight their advantages and their limits.

One thing in common: both technologies use Machine Learning algorithms. But there are several major differences:

– Deep Learning produces neural networks that are “black box” models, while the xtractis AI produces “white box” decision systems based on fuzzy rules, understandable by the business expert.

– A neural network is a global model: any modification of parameters changes the behavior of the decision-making system; while the xtractis model is composed of local sub-models (fuzzy rules): the modification of the premise (IF… part)or of the conclusion (THEN… part) of a rule has an impact only in the local control area of the rule. xtractis can thus improve the global model locally in order to gain efficiency while preserving the explainability.

– Unlike a neural network, it is easy to demonstrate that the output of a fuzzy rule-based system will vary continuously and gradually between a lower bound and an upper bound at any point in the decision space. The stability of such a system is thus proven, especially for critical decisions, which allows its certification.

– To learn, Deep Learning requires a very large amount of data, while the xtractis AI can handle small datasets.

– Deep Learning, like any statistical approach, handles missing data by imputation (assignment of an estimated value), which introduces a bias in the data before processing; while the xtractis AI preserves this state of ignorance (lack of information) by supposing that all the values of the unspecified variable are possible (extremal level of fuzziness).

– Deep learning algorithms are usually fixed by the modeler, while the xtractis AI is able to improve its own inductive learning strategies to enhance its performance continuously.

– Deep Learning uses the unique operators of binary logic, tensoral calculus and probability measure, while the xtractis AI uses an infinite number of multivalent logical operators, an infinite number of compositional operators by relational anchoring and an infinite number of generalized measures of possibility and necessity, which gives it greater degrees of freedom in non-linear modeling.

– The algorithms of Deep Leaning are open source, while those of xtractis are proprietary.

The decision system created by an AI must be explainable in all areas with high societal impact: Smart Industry, Transportation, Defense, Security, Nuclear, Finance, Health, Law, Politics. Indeed, such systems must be audited and certified by the regulator before their deployment.

However, a “black box” system could be used for Marketing or CRM applications such as product or service recommendation.

For example, a French bank could use neural networks to create chatbots, but will have to rely on explainable modeling (analytic regression, rule based-system, decision tree) to model the behavior of its clients or to evaluate the score of a credit file. The first model will be audited by the Marketing Department, while the second one will be audited and certified by the ACPR and the ECB.

In fact, since the entry into force of the European General Data Protection Regulation (GDPR), the “right to explanation” granted to any natural or legal person, undergoing an automated decision imposes on any designer of such an automated decision-making system to justify the rules that produced the decision. This prohibits de facto any “black box” decision system.