Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

A more selective integration function to improve deep neural network models of visual perception

Michael W. Spratling¹, Heiko H. Schütt²; ¹University of Luxembourg, ²University of Luxemburg

Presenter: Michael W. Spratling

Human visual perception remains substantially more robust than computer vision. We hypothesised that this might be due to the higher selectivity of biological neurons compared to artificial neurons which typically employ a linear integration function that is poor at feature detection. To test this hypothesis we replaced the convolutional layers in deep neural networks (DNNs) with a new integration function, the Consistent Intensity Metric (CIM). We trained networks based on CIM on six benchmark image classification tasks (MNIST, FashionMNIST, SVHN, CIFAR10, CIFAR100, and TinyImageNet) and compared the performance of these networks with equivalent convolutional neural networks matched to have an equal number of parameters. Consistent with our hypothesis, the CIM-based networks were better able to generalise from the training data. This was demonstrated by higher accuracy on both the standard test data and distorted input images (the common corruptions data-sets). Furthermore, test images that did not belong to any of the categories in the training data-set were less likely to be misclassified as belonging to one of the known categories. Our results suggest that using a more selective integration function can help address some of the reliability and robustness issues of DNNs. As these issues do not affect humans, this modification also makes DNNs functionally more similar to the biological visual system.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF