MAIA, M. G.; PEREIRA, EANES TORRES.; http://lattes.cnpq.br/5915606893493967; MAIA, Matheus Gomes.
Resumen:
Deep neural networks have enabled remarkable advances in
image processing and analysis. The complexity resulting from the adoption of these
Machine learning models, characterized by an increasing number of parameters, induce knowledge representations and decision flows that go beyond
human understanding, especially when it comes to models used in
visual tasks such as Convolutional Neural Networks. This research reviews
and structures approaches and interpretability techniques, which aim to expose, in a
understandable way, the workings or internal knowledge of these models
convolutional. In this research, the objectives of the
interpretability techniques and the characteristics of revised interpretable models
From. The recent scientific debate on the evaluation of new interpretability techniques was presented and explored in a practical way in a case study. O
This study was carried out in an unusual image niche for research that evaluates
interpretability techniques, that of medical diagnosis, which proved to be challenging
and rich in learning. The case study considers, end-to-end, the steps
of training, interpretability and evaluation of the techniques, something that can be easily reproduced in other data sets. For the training stage
two convolutional architectures were used trained on two composite sets.
by images of medical exams, being these, thoracic X-ray and tomography of
optical coherence (OCT). For the explanation step, ten interpretability techniques were used to produce explanations for the trained models. Finally,
for the technique evaluation stage, the explanations were submitted to three evaluations: Guide to Disorders, Randomization of Labels and Pointing Game. Each
evaluation brought challenges and lessons learned about its approaches and necessary resources. For example, the Guide to Disturbances assessment favors concise,
but is hampered by an effect known as Şout of distribution entries,
in addition to having a high processing cost. This dissertation aims to
main contributions to the production of a narrative and propositional bibliographic review on the area of interpretability of neural networks, including the discussion of
topics such as taxonomy of approaches to interpretability and categorization
of quantitative assessments of attribution techniques, in addition to a case study
on evaluations of techniques that interpret Convolutional Neural Networks.
