Hierarchical feature representation reduces the Müller-Lyer effect A Zeman1, K Brooks1, O Obst2 |
|---|
1Department of Psychology, Macquarie University, Australia |
| Deep neural networks inspired by visual cortex demonstrate superior pattern recognition compared to their shallow counterparts. These artificial neural networks (ANNs) with hierarchical feature representation also provide a new method for investigating visual illusions. Recently, a state-of-the-art computational model of biological object recognition, HMAX, was found to exhibit a bias in line length of when shown Müller-Lyer stimuli. The Müller-Lyer Illusion (MLI) is a visual illusion wherein a line appears elongated with arrowtails and contracted with arrowheads. The combined and separate contributions of training stimuli and elements of neural computation can be explored in ANNs to investigate the possible causes of an illusory effect. In this study we investigate whether the MLI occurs because of feature representation built from “simple” and “complex” cells or whether using an SVM as the decision-making module drives the effect. We ran dual category line length discrimination experiments in both the full HMAX model (including an SVM stage) and an SVM-only model. Unexpectedly, the SVM demonstrated an even larger misclassification of line length than shown by HMAX. These results indicate that a simple-complex neural architecture is not necessary to simulate the illusion but rather suggests that filtering and max pooling operations reduce the Müller-Lyer effect. |
| Up Home |
|---|