On the Design, Implementation and Application of Novel Multi-disciplinary Techniques for explaining Artificial Intelligence Models

  1. Barredo Arrieta, Alejandro
Dirigida por:
  1. Javier del Ser Lorente Director
  2. Sergio Gil López Director/a

Universidad de defensa: Universidad del País Vasco - Euskal Herriko Unibertsitatea

Fecha de defensa: 28 de enero de 2022

Tribunal:
  1. David Camacho Fernández Presidente/a
  2. Manuel María Velez Elordi Secretario
  3. Natalia Díaz Rodríguez Vocal

Tipo: Tesis

Teseo: 156942 DIALNET lock_openADDI editor

Resumen

Artificial Intelligence is a non-stopping field of research that has experienced some incredible growth lastdecades. Some of the reasons for this apparently exponential growth are the improvements incomputational power, sensing capabilities and data storage which results in a huge increment on dataavailability. However, this growth has been mostly led by a performance-based mindset that has pushedmodels towards a black-box nature. The performance prowess of these methods along with the risingdemand for their implementation has triggered the birth of a new research field. Explainable ArtificialIntelligence. As any new field, XAI falls short in cohesiveness. Added the consequences of dealing withconcepts that are not from natural sciences (explanations) the tumultuous scene is palpable. This thesiscontributes to the field from two different perspectives. A theoretical one and a practical one. The formeris based on a profound literature review that resulted in two main contributions: 1) the proposition of anew definition for Explainable Artificial Intelligence and 2) the creation of a new taxonomy for the field.The latter is composed of two XAI frameworks that accommodate in some of the raging gaps found field,namely: 1) XAI framework for Echo State Networks and 2) XAI framework for the generation ofcounterfactual. The first accounts for the gap concerning Randomized neural networks since they havenever been considered within the field of XAI. Unfortunately, choosing the right parameters to initializethese reservoirs falls a bit on the side of luck and past experience of the scientist and less on that of soundreasoning. The current approach for assessing whether a reservoir is suited for a particular task is toobserve if it yields accurate results, either by handcrafting the values of the reservoir parameters or byautomating their configuration via an external optimizer. All in all, this poses tough questions to addresswhen developing an ESN for a certain application, since knowing whether the created structure is optimalfor the problem at hand is not possible without actually training it. However, some of the main concernsfor not pursuing their application is related to the mistrust generated by their black-box" nature. Thesecond presents a new paradigm to treat counterfactual generation. Among the alternatives to reach auniversal understanding of model explanations, counterfactual examples is arguably the one that bestconforms to human understanding principles when faced with unknown phenomena. Indeed, discerningwhat would happen should the initial conditions differ in a plausible fashion is a mechanism oftenadopted by human when attempting at understanding any unknown. The search for counterfactualsproposed in this thesis is governed by three different objectives. Opposed to the classical approach inwhich counterfactuals are just generated following a minimum distance approach of some type, thisframework allows for an in-depth analysis of a target model by means of counterfactuals responding to:Adversarial Power, Plausibility and Change Intensity.