Research projects

Artificial Intelligence (AI) has achieved remarkable levels of effectiveness in many domains. Examples are the use of chatbots for customer support; medical imaging to help doctors diagnose medical conditions or autonomous vehicles to assist drivers.

AI systems face common challenges in their development: they require a different and specialized skillset; they are hard to specify, test, verify and debug; and are complex to evolve and maintain. Explainability (the ability to justify the decisions made by AI components) and AI ethics (the ability to define and check ethical principles to be respected in such decisions, e.g. the lack of bias) are also of special importance.

Moreover, AI systems are almost always part of a larger software system that embodies them. This combination is usually referred to as AI-enhanced software or simply smart software. Currently, the AI part of smart software is developed separately from the rest. This poses additional challenges: defining the communication between the AI elements and the traditional ones, the end-to-end testing of the global system, their co-evolution, … This complexity can further increase the AI-divide between the tech giants and the rest of the World.

This research project aims to change this situation. Our goal is to simplify the specification, generation, testing, deployment and evolution of any type of smart software thanks to a LOw-COde development platform for Smart Software (LOCOSS).

Low-code application platforms accelerate software delivery by dramatically reducing the amount of hand-coding required. Low-code can be seen as a specific style of Model-Driven Engineering (MDE), a software development paradigm where models rather than source code are the core asset. Models provide an abstract and simplified view of a software system focused on a specific perspective. MDE raises the level of abstraction in software engineering with the benefits of higher quality, technology independence and reduced development costs.

As such, we propose to bring the power of low-code to the development of smart software. This combination has not been deeply studied so far. We believe it can disrupt how smart software is built, lowering the barrier entry to AI development, improving its quality and reducing the overall development effort.

To achieve this ambitious goal, the project will pursue these key research contributions:

• A set of domain-specific languages to facilitate the specification of AI components (e.g. intelligent conversational interfaces, recommender services, …) and their interaction with the non-AI ones.

• A model-based approach for training and optimizing the AI components as part of the smart software specification.

• Explicit high-level formalization of complex non-functional requirements such as security, privacy or fairness, which may otherwise become implicit or scattered in the implementation.

• Code-generation techniques to implement the AI components on top of state-of-the-art AI libraries isolating as much as possible the AI designer from low-level technical details.

• Novel verification, validation and testing techniques to include the quality evaluation of AI components.

The results of this project will have a significant technical, economic and social impact by expanding the number of potential smart software developers and reducing the time-to-market for this type of software, improving the competitiveness of Spanish companies.