In Late-Medieval panel paintings from the Tuscan area, mechanical tools called punches were used to impress repeated motifs on gold foils to create decorative patterns. Such patterns can be used as clues to objectively support the attribution of the paintings, as proposed by art historian Erling S. Skaug in his decades-long study on punches. We investigate the feasibility of employing automatic pattern recognition techniques for accelerating the process of classification of punches by experts working in the field. We propose a system composed of (a) a Convolutional Neural Network for categorizing a punch contained in a frame, and (b) an additional component for uncertainty estimation, aimed at recognizing possible Out-of-Distribution (OOD) samples. After collecting a set of 3815 punches from four 14 th century panel paintings from Tuscany, we train a Convolutional Neural Network which achieves very high test-set accuracy. As far as the uncertainty estimation is concerned, we experiment with two techniques, OpenGAN and II-loss, both exhibiting very positive results. The former seems to work better on specific data extracted from images of panel paintings, while the latter showcases a more consistent behavior when considering additional OOD data obtained randomly. These outcomes indicate that an application of our system in support of experts is feasible, although we subsequently show that additional experiments on larger datasets might be required.
3 February 2022 in The Knowledge Engineering Review
Artificial neural networks (ANNs) can be employed as controllers for robotic agents. Their structure is often complex, with many neurons and connections, especially when the robots have many sensors and actuators distributed across their bodies and/or when high expressive power is desirable. Pruning (removing neurons or connections) reduces the complexity of the ANN, thus increasing its energy efficiency, and has been reported to improve the generalization capability, in some cases. In addition, it is well-known that pruning in biological neural networks plays a fundamental role in the development of brains and their ability to learn. In this study, we consider the evolutionary optimization of neural controllers for the case study of Voxel-based soft robots, a kind of modular, bio-inspired soft robots, applying pruning during fitness evaluation. For a locomotion task, and for centralized as well as distributed controllers, we experimentally characterize the effect of different forms of pruning on after-pruning effectiveness, life-long effectiveness, adaptability to new terrains, and behavior. We find that incorporating some forms of pruning in neuroevolution leads to almost equally effective controllers as those evolved without pruning, with the benefit of higher robustness to pruning. We also observe occasional improvements in generalization ability.