A precursory insight into the behavior of CNNs to analyze visual similarities of sketches | ||||
| Egyptian Journal of Pure and Applied Science | ||||
| Article 1, Volume 59, Issue 2, December 2021, Page 1-14 PDF (2 MB) | ||||
| Document Type: Original Article | ||||
| DOI: 10.21608/ejaps.2021.97317.1007 | ||||
 View on SCiNiTO | ||||
| Authors | ||||
Nohayr Abdelmoneam ; Soheir Khamis ; Ahmed Abdelfattah 
| ||||
| Mathematics Department, Faculty of Science, Ain Shams University, Cairo, Egypt | ||||
| Abstract | ||||
| Artificial neural networks (ANNs) have been showing a great performance in artificial intelligence (AI)-related tasks in many fields for the last couple of decades or so. Nevertheless, there is also still an apparent lack of any clear understanding of its learning behavior, which is triggering the research of ANNs interpretation. In this article, we put forward a new technique to investigate a specific learning behavior of Convolutional Neural Networks (CNNs) using the problem of free-hand sketch classification as an interesting testbed. Our aim is to look into possible factors that contribute to the classification process and transfer of learning between different categories. We designed and analyzed two advanced experiments to deduce possible learning attributes that could be shared by both CNNs and humans. Our preliminary results show a huge potential of CNNs to detect visual similarities among hand-drawn sketches of different object categories, which seems to resemble that of humans to some extent. | ||||
| Keywords | ||||
| Feature visualization; CNN interpretation; sketch recognition; visual similarities; learning transfer | ||||
|
Statistics Article View: 324 PDF Download: 252 |
||||
