Early neural networks were inspired by biology: the McCulloch-Pitts neuron, the Perceptron and the Neocognitron were all attempting to imitate the functioning of the brain. [7] To explore ways in which the study of neurology can aid the development of Artificial Intelligence solutions, we describe the research “On-Off Center-Surround Receptive Fields for Accurate and Robust Image Classification” (by Babaiee et.al.[12]), in which a Convolutional Neural Network (CNN) is improved with kernels inspired by the On-Off-Center-Surround receptive fields of the vertebrate retina. Finally, we illustrate our experiments, devised to expand on Babaiee’s research. With the premise that kernels inspired by the retina were able to improve the performance of the algorithm, we implemented new kernels inspired by the early visual system. Specifically, we made kernels that reproduce the function of simple cells in the area V1 of the visual cortex, and kernels that try to replicate some functions of the complex cells of visual area V2. Then we tested several CNNs that implement these kernels. These networks did not achieve a better performance than Babayee’s CNN, but some did improve upon the CNN used as control. However, our tests were limited in scope and hence the basic idea may still be implemented successfully. We will therefore describe possible solutions that could be tested in future research.
Augmenting convolutional neural networks with kernels inspired by the early visual system
ROVOLETTO, MATTEO BRUNO
2021/2022
Abstract
Early neural networks were inspired by biology: the McCulloch-Pitts neuron, the Perceptron and the Neocognitron were all attempting to imitate the functioning of the brain. [7] To explore ways in which the study of neurology can aid the development of Artificial Intelligence solutions, we describe the research “On-Off Center-Surround Receptive Fields for Accurate and Robust Image Classification” (by Babaiee et.al.[12]), in which a Convolutional Neural Network (CNN) is improved with kernels inspired by the On-Off-Center-Surround receptive fields of the vertebrate retina. Finally, we illustrate our experiments, devised to expand on Babaiee’s research. With the premise that kernels inspired by the retina were able to improve the performance of the algorithm, we implemented new kernels inspired by the early visual system. Specifically, we made kernels that reproduce the function of simple cells in the area V1 of the visual cortex, and kernels that try to replicate some functions of the complex cells of visual area V2. Then we tested several CNNs that implement these kernels. These networks did not achieve a better performance than Babayee’s CNN, but some did improve upon the CNN used as control. However, our tests were limited in scope and hence the basic idea may still be implemented successfully. We will therefore describe possible solutions that could be tested in future research.File | Dimensione | Formato | |
---|---|---|---|
Tesi Matteo Bruno Rovoletto.pdf
accesso aperto
Dimensione
2.74 MB
Formato
Adobe PDF
|
2.74 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/30217