Biophysical neural adaptation mechanisms enable deep learning models to capture dynamic retinal computation

Abstract

Neural adaptation is a universal feature of neural systems that modulates the output based on input conditions, enabling efficient encoding of sensory inputs without saturation or loss. In contrast, conventional artificial neural networks (ANNs) lack these adaptational mechanisms, resulting in inaccurate neural predictions under changing input conditions. Can embedding neural adaptive mechanisms in ANNs improve their performance and generate biologically-plausible models? To address this question, we introduce a new type of convolutional neural network (CNN) layer incorporating photoreceptor biophysics and adaptation mechanisms to model light-dependent photoreceptor sensitivity and kinetics. Under changing input light conditions, CNNs that include the new photoreceptor layer integrated as a front-end perform better than conventional CNN models at predicting: (1) monkey and rat retinal ganglion cell responses; and (2) context-dependent changes in neural sensitivity. This study demonstrates the potential of embedding neural adaptive mechanisms in deep learning models, enabling them to adapt dynamically to evolving inputs.