Towards an “Innate Learning” Efficient Coding Model using Spontaneous Neural Activity
DescriptionTraining computer vision systems is often driven by available data, however, the human visual system develops without the benefit of stored data. Here we demonstrate how using a simplified computational model. Visual development progresses from an innate stage to visual-based experiential learning, where the former is initially driven by chemical axonal guidance cues and the latter through direct visual experience. Adaptation to visual experience after eye-opening has been thoroughly observed. Our visual cortex appears to also adapt to internally-generated, spontaneous patterns of neural activity before eye-opening using a process similar to what is observed with visual stimuli after eye-opening - a process we refer to as “innate learning”. For adults, an efficient coding approach has been applied to natural scenes to model early visual processing using sparse and independent coding objectives that mimic neural processing constraints. For visual development prior to eye-opening, we demonstrate similar training with spontaneous activity patterns which contain the same low-level statistical structure. In this work, we use a model of retinal waves prior to eye-opening as stimuli for efficient coding. The produced efficient coding filters resemble neural receptive fields suggesting the same efficient coding principle can be used both before and after eye-opening. For future work, the integration of spontaneous patterns of neural activity can be valuable for model training, generalization, and refinement. Such unsupervised pre-training may reduce training time in object recognition, but the implications of integrating spontaneous activity in later layers may also improve generalization and is worthy of further study.