I spy with my little eye... But what can rats spy? Investigating the rodent model for high-level visual processing

I spy with my little eye, a well-loved childhood game that many people like to play. As scientists, we were wondering how the brain solves such a task, where certain objects have to be found in a scene, or known objects have to be recognized from different viewpoints. This is exactly what the domain of invariant object recognition focuses on. To understand how the human vision system performs this task, we have to take a closer look at the neural underpinnings of the human brain. From previous studies, we know that the brain processes visual information through two functionally and anatomically distinct pathways. The pathway we focused on in this dissertation is the ventral visual pathway, or the what stream, as it focuses on object identity and recognition. Visual information consists of two different kinds of features: low-level features and high-level features and these features are processed differently in the human brain.

In this dissertation, we wanted to investigate this ventral visual stream by using two models with which we can or try to model high-level visual processing.

The first model is the rat model, and we have used this animal model throughout all three studies. In our first two studies, the animals were trained in a face categorization task, using stimuli of human faces as targets and random objects (Chapter 2) or noisy stimuli (Chapter 3) as distractors. We found that rats were capable of learning these complex categorization tasks and that they were able to generalize to new, unseen stimuli, even after applying transformations to these stimuli. In our second study, we focused on contrast features, a low-level feature that has been investigated in monkeys and humans in the domain of face detection. We found that rats use contrast features and furthermore, we found similarities between human and rodent vision.

The second model that we used are computational models where we simulate object recognition in convolutional deep neural networks (DNNs). We investigated the properties of these networks such that it simulates the recognition based on contrast features that we find in rats and humans. In our last study (Chapter 5), we used a DNN to select stimuli that target different visual strategies. We trained rats and humans in a two-dimensional categorization task to investigate the benefit of using a computational approach for designing a visual task and selecting stimuli. We found interesting opposite results between rat and human performance in terms of their visual strategies, and thus evidence for the additional value of using a computational approach in vision studies.

Our findings of these three studies suggest that the rat model can be, to some extent, used for high-level visual processing. Future directions are discussed in terms of adding neural data collection, as well as focusing on the ecological validity of the visual tasks.