Affordable Artificial Intelligence – AffAI

The Affordable AI (AffAI) proposal aims to address three bottlenecks preventing industrial end users from tapping into the potential of AI technologies:

* a considerable know-how is required to build data-driven intelligent systems. The exploration of the processing pipelines (pre-processing and learning algorithms, and their hyper-parameters), referred to as AutoML, is increasingly acknowledged as non sustainable due to its huge computational cost.

* black-box models learned from data are not satisfactory for reliability, fairness, and accountability reasons.

* in many industrial sectors one can only gather small data (contrasting with the wealth of data behind the AI breakthroughs). One must thus find a way to leverage the knowledge of the domain experts to palliate the lack of data.

AffAI builds up a consortium involving 2 international level Universities, renowned for their leadership in AI: Univ. Paris-Saclay and TU Dortmund University, and 2 AI SMEs, RapidMiner and MyData Models.

This consortium is able to: i) actually understand the main needs of industrial end users; ii) design accurate solutions by leveraging and adapting the best state of the art; iii) implement these solutions in a professional and efficient way; iv) enforce their fast dissemination thanks to an active network of corporate customers and a large size user community.

The above three bottlenecks will be addressed by, building upon the complementary expertises of the consortium:

1. AutoML will be extended to accommodate the fact that end users are actually interested in several performance indicators (including the overall computational cost), making the search for a good pipeline a multi-criteria optimization problem.
A most innovative aspect is to address Embeddable AutoML, where the training of the intelligent systems is done on-board, subject to specific resource constraints.

2. Two functionalities will be provided to open the black-box of the learned intelligent systems: a sensitivity analysis, showing how the current decision proposed by the system depends on the case and its context; a causal model, characterizing how the features of the case depend on each other. This causal model will support the exploration of *interventions*, aimed to modify the features when possible in order to bring desired effects.

3. The shortage of data, and possibly the partially incomplete specifications of the sought intelligent systems will be addressed by leveraging the expertise of the domain expert, through interaction. On the one hand, the shortage of data will be palliated by active learning (asking the domain expert to label a few dozens most informative cases). On the other hand, the models compliant with the evidence will iteratively be demonstrated to the domain expert, who will be able to provide a simple feedback about which model is more appropriate than the other. This interactive dialog will support the gradual learning of the expert's preferences and expectations, and will be used to shape the behavior of the eventual intelligent system.


The validation and transfer of the proposed functionalities will be achieved along an original 2-tier mechanism:
* The two SMEs will provide 2 representative use cases, respectively related to predictive maintenance and process optimization, on which they can gather and share data.
* The proposed functionalities will be first validated on these use cases.
* In a third phase, these functionalities will be applied on similar but confidential industrial applications. The SMEs will thus be able to assess and enforce the robustness and generality of these functionalities, interacting back and forth with the Labs and with their end-users.

Ultimately, AffAI is relevant to Green AI, as it expects to significantly decrease the computational cost of building intelligent systems, and to the Democratisation of AI, through enabling industrial end users (and others) to build and trust intelligent systems.