Currently, in contrast to shallow models exploited in the past, most deep learning systems extract features automatically, and to do that, they tend to rely on a huge amount of labeled data. Whereas the quality of dataset used to train neural networks has a huge impact on the models’ performance, those datasets are often noisy, biased and sometimes even contain incorrectly labeled samples. Moreover, deep neural networks (DNNs) are black-box models that usually have tens of layers with millions of parameters, and very complex latent space, which make their decisions very hard to interpret. Such fragile models are increasingly used to solve very sensitive and critical tasks. Therefore, the demand for a clear reasoning and correct decision is very high, especially when DNNs are used in transportation (autonomous cars), in healthcare, for legal systems, finances, and military.
To address those challenges the project aims to develop methods of Explainable Artificial Intelligence (XAI) which might help to uncover and reduce the problem of bias in data. The project involves investigation and integration of explainability into new and existing Artificial Intelligence systems, and mostly focuses on Deep Neural Networks in the field of Computer Vision. One of the ways of categorizing XAI methods is to divide them into local and global explanations. Local analysis aims to explain a single prediction of a model, whereas a global one tries to explain how the whole model works in general. The project aims to develop novel methods of both local and global explainability to help explain deep neural network-based systems in order to justify them, to control their reasoning process, and to discover new knowledge.
Deaf people are affected by many forms of exclusion, especially now in the pandemic world. HearAI aims to build a deep learning solution to make the world more accessible for the Deaf community and increase the existing knowledge base in using AI for Polish Sign Language.
In the last twenty years the interest of automated skin lesion classification dynamically increased partially because of public datasets appearing. Automated computer-aided skin cancer detection in dermatoscopic images is a very challenging task due to uneven datasets sizes, the huge intra-class variation with small interclass variation, and numerous artifacts. During my work on the project I approached the problem in two ways:
with hand-crafted features based on extended ABCD rule and a shallow neural network, with Convolutional Neural Networks.