Lie symmetries in cortical inspired CNNs

Bertoni, Federico (2022) Lie symmetries in cortical inspired CNNs, [Dissertation thesis], Alma Mater Studiorum Università di Bologna. Dottorato di ricerca in Matematica, 35 Ciclo. DOI 10.48676/unibo/amsdottorato/10098.
Documenti full-text disponibili:
[img] Documento PDF (English) - Richiede un lettore di PDF come Xpdf o Adobe Acrobat Reader
Disponibile con Licenza: Creative Commons Attribution Non-commercial ShareAlike 4.0 (CC BY-NC-SA 4.0) .
Download (20MB)

Abstract

Our scope in this thesis is to propose architectures of CNNs in such a way to model the early visual pathway, including the Lateral Geniculate Nucleus and the Horizontal Connectivity of the primary visual cortex. Moreover, we will show how cortically inspired architectures allow to perform contrast perceptual invariance as well as grouping and the emergence of visual percepts. Particularly, the LGN is modeled with a first layer l0 containing a single filter Ψ0 that pre-filters the image I. Since the RPs of the LGN cells can be modeled as a LoG, we expect to obtain a radially symmetric filter with a similar shape; to this end, we prove the rotational invariance of Ψ0 and we study the influence of this filter to the subsequent layer. Indeed, we compare the statistic distribution of the filters in the second layer l1 of our architecture with the statistic distribution of the RPs of V1 cells of a macaque. Then, we model the horizontal connectivity of V1 implementing a transition kernel K1 to the layer l1. In this setting, we study the vector fields and the association fields induced by the connectivity kernel K1. To this end, we first approximate the filters bank in l1 with a Gabor function and use the parameters just found to re-parameterize the kernel. Thanks to this step, the kernel is now re-parameterized into a sub-Riemmanian space R2 × S1. Now we are able to compare the vector and association fields induced by K1 with the models of the horizontal connectivity.

Abstract
Tipologia del documento
Tesi di dottorato
Autore
Bertoni, Federico
Supervisore
Dottorato di ricerca
Ciclo
35
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
Lie symmetries ; CNN architectures; Lateral Geniculate Nucleus ; Primary Visual Cortex ; Retinex theory ; Emergence of visual percepts
URN:NBN
DOI
10.48676/unibo/amsdottorato/10098
Data di discussione
17 Febbraio 2022
URI

Altri metadati

Statistica sui download

Gestione del documento: Visualizza la tesi

^