RRR-XAI · Right for the Right Reason eXplainable Artificial Intelligence
„Хоризонт Европа“ — Действия „Мария Склодовска-Кюри“
- Период
- 2022-11-01 → 2024-10-31
- Финансиране от ЕС
- 165 313 €
- Участници
- 2
- Схема
- HORIZON-TMA-MSCA-PF-EF
Линиите свързват координатора с партньорите.
Накратко на български
Изкуственият интелект се анализира, за да се разбере защо невронната мрежа свързва конкретен вход с определен резултат. Това помага да се открият грешки в данните и да се обяснят логическите причини за решенията на машината на разбираем език.
Кратко обяснение, генерирано от езиков модел по текста на CORDIS. Оригиналът е по-долу.
Резултати накратко
RRR-XAI: Right for the Right Reason eXplainable Artificial Intelligence
The overall purpose of RRR-XAI was making deep learning (DL) explainable under the right for the right reasons (RRR) philosophy, by creating explanations endorsed by the reasoning of a domain expert, using the X-NeSyL (eXplainable NeSy Learning) methodology. To achieve this I followed the rationale behind XAI under the RRR philosophy. First, performing analyses to understand two types of phenomena that cause trouble in DNNs. The second part consists of using NeSy computation to communicate such phenomena. Research and innovation objectives included: Objective 1.1: To understand the phenomenon: Making DL explainable consists of being able to pinpoint why a deep neural network (DNN) associates an output to a given input. Could we use inherent properties of data to diagnose why a DNN produces an output y for a given input x? The aim is to design algorithms based on instance quality measures of data, as a proxy to guarantee transparency, to get understanding of such mechanisms. The hypothesis of O1.1 to test is that data preprocessing procedures and intrinsic aspects of data may highlight the provenance of bias or learning tricks often abused by DL models, and these can be measured with eXplainable AI (XAI) metrics. Objective 1.2: To communicate the phenomenon: Once identified the root cause leading a DNN to associate an output to a given input, What if we could interrogate it, or “let the DNN talk”? The objective is being able to convey the constraint that led the DNN to such input-output association to different audiences, using high-level (symbolic, relational) concepts in natural language. The hypothesis here is that 1) The natural language explanation (NLE) can synthesize a formal (logical, causal or counterfactual) explanation while sacrificing little-to-no performance; 2) This NLE can be accurate enough to correct the model's critical and unfair errors.
Текст от CORDIS, на английски · Данни: CORDIS, © Европейски съюз
Цел на проекта
Deep Deep Learning (DL) is a form of machine learning (ML) that enables computers to learn from experience and understand the world in terms of a hierarchy of concepts. This hierarchy allows a DL model to learn complicated concepts by building them out of simpler ones. A graph of these hierarchies would be many layers deep, and thus its name. A Deep neural network (DNN) is based on an artificial neural network model, and its core strength is that there is no need for human assistance to formally specify all the knowledge that the model needs. This makes DNNs represent the state of the art in Artificial Intelligence (AI). Despite their top performance and ubiquity of applications (from Healthcare to autonomous cars), DNNs suffer serious shortcomings. First, DNNs are considered black box models, i.e., with complex and opaque algorithms, hard to interpret and diagnose. Second, they suffer from bias, and the testing protocols for automatic recognition are not fair, as they learn patterns in the data that are not correlated to the output; e.g. they may focus on areas outside the lung in X ray images to predict the presence of COVID-19. Although DNNs outperform many other methods, they often are not right for the right reasons (RRR). RRR-XAI tackles this mismatch and bridges this gap through a tight integration of DL and symbolic AI, with the principal objective of making DL explainable. To achieve this I will follow the rationale behind XAI under the RRR philosophy and perform analyses to understand two types of phenomena that cause trouble in DNNs. Second, I will use: 1) Domain knowledge expertise as supporting evidence to explain a particular model output; 2) Neural-Symbolic computation to communicate the explanation of such phenomena in natural language. I will study two practical use cases where supporting explanations of the model output are critical: a) COVID-19 prediction from chest X-Ray images, and b) Weapon detection in alarm systems and crowds from images.
Оригинален текст от CORDIS (на английски).
Участници
Връзки
Данни: CORDIS, © Европейски съюз