Pre-training artificial neural networks with spontaneous retinal activity improves image prediction

The processing of natural visual environments rich with motion is central to animal survival. Many aspects of perceiving and detecting visual motion are already mature before the onset of visual experience. How the visual system develops to enable these functions in the absence of external stimuli remains unclear. A prominent transient feature of developing circuits is the ability to generate complex spatiotemporal patterns of spontaneous activity, even before sensory organs mature. For example, before vision onset in the mammalian retina, activity propagates in the form of spatiotemporally structured patterns known as waves, which have been implicated in the refinement of different aspects of circuit connectivity and function. This includes the formation of gross retinotopic maps and the refinement of receptive fields in downstream visual areas like the thalamus, superior colliculus, and visual cortex. Recent work showed that retinal waves in mice contain statistical features of real-world visual stimuli such as optic flow (Ge et al., 2021), suggesting they could also play an important role in preparing the visual system for motion processing.

While recent biologically-constrained networks implicate retinal waves in network connectivity refinements, we wondered whether state-of-the-art artificial neural network (ANN) models trained on natural movies show improved performance if pre-trained with retinal waves. Indeed, we found that pre-training ANNs with retinal waves significantly improves the processing of real-world visual stimuli and temporarily accelerates learning. We further investigated how other aspects of retinal waves, such as their propagation bias, may influence the learning speed and network performance on both retinal wave and natural movie datasets. Overall, our work sheds light on the functional role of spatiotemporally patterned spontaneous activity in the processing of motion in natural scenes, suggesting that they provide a training signal to prepare the developing visual system for adult visual processing.