Mapping Visual Receptive Fields in Convolutional Neural Networks

Abstract

Receptive field (RF) mapping is a fundamental technique for analyzing visual neurons. Neurophysiologists often use simple stimuli such as light spots or bars to locate visual RFs. While this approach is valid for early visual areas, where neurons are selective for simple features, its effectiveness in higher areas is less certain. The continued use of this technique for mapping RFs in higher areas may lead to misinterpretations of RF properties. With the advent of deep neural networks, we now have models that allow for the extraction of ground-truth RF properties. This methodological thesis focuses on developing and accurately quantifying RF mapping techniques in deep neural networks. Our findings indicate that RF mapping using bars is unreliable beyond early-to-mid layers, as bars do not drive neurons as effectively as natural images. This underscores the need for more robust RF mapping approaches for higher layers in deep neural networks but also for neurophysiology in general.