Benchmarking birds’ core object recognition abilities with rodents and primates

Jonas Rose – Ruhr Universität Bochum

The primate visual system can recognize objects despite large changes in their appearance on the retina (e.g., size, position, lighting etc.). This ability is known as invariant object recognition, and it is the foundation for cognitive and memory processes that depend on visual information. Despite invariant object recognitions importance in vision, its computational mechanisms remain poorly understood. Primates have traditionally been the animal model of choice for investigations into invariant object recognition and all the species that have been tested display advanced object invariance capabilities. These capabilities depend on a hierarchical stage of processing equivalent to inferior temporal cortex. More recently, rodents have been shown to also be capable of object invariance, though at a lower-level equivalent to mid-level primate visual areas (V4). An unsolved question is whether birds – who have advanced eyesight and visual systems like primates – are also capable of invariant object recognition despite having a very different neuronal architecture to the mammalian brain. In the current application, we will benchmark the performance of pigeons and jackdaws on the same object invariance tasks as what have been used to evaluate primates and rodent performance. Corvids display advanced cognitive abilities like primates – such as face-recognition, tool use etc. – and have a greatly expanded endbrain relative to less cognitively advanced birds like pigeons. We will determine for the first time if the expanded jackdaw brain is capable of supporting object invariance at the level of primate inferior temporal cortex. By benchmarking the performance of birds in directly comparable object invariance tasks, we will understand what the universal and essential computations are that mediate object invariance across brains with very different architectures.These investigations are broken down into three stages. Firstly, we will train pigeons and jackdaws on a simple object discrimination task to evaluate if their visual systems can contend with relatively simple object transformations. Secondly, we will train pigeons and jackdaws on a complex object discrimination task to determine if the avian brain can contend with complex transformations of objects in complex backgrounds, revealing the upper limit of shape complexity that pigeons and jackdaws can compute from images. Lastly, we will perform large-scale recordings from different stages of the visual system in pigeons and jackdaws while they passively view images of objects. The final stage enables us to determine how the avian nuclear neuronal architecture represents the physical structure of objects. These investigations - in collaboration with the Issa lab at Columbia University - will reveal the computational underpinnings of invariant object recognition. These insights will be used to help guide the development of the next generation of machine vision systems.