Fig. 1: A deep neural network model accurately predicts mouse V1 responses to natural scenes.
From: Functional bipartite invariance in mouse primary visual cortex receptive fields
a, Schematic of the optimization of MEIs and VEIs. The vertical axes depict the activation of two model neurons as a function of two example image features. Left, neuron without obvious invariance; right, neuron with phase invariance to its optimal stimulus. Black curves illustrate optimization trajectories for MEI from different initializations (left) and for VEIs as perturbations starting from the MEI along the invariance ridge (right). b, Schematic of the inception loop paradigm. On day 1, we presented sequences of natural images and recorded in vivo neuronal activity using two-photon calcium imaging. Overnight, we trained an ensemble of CNNs to reproduce the measured neuronal responses and synthesized artificial stimuli for each target neuron in silico. On day 2, these stimuli were presented to the same neurons in vivo to compare measured and predicted responses. c, We presented 5,100 unique natural images to an awake mouse for 500 ms each, interleaved with gray screen gaps of random length between 300 and 500 ms. A subset of 100 images was repeated ten times to estimate neuronal response reliability. Neuronal activity in V1 L2/3 was recorded at 8 Hz using wide-field two-photon microscopy. Behavioral traces including pupil dilation and locomotion velocity were also recorded. d, CNN model architecture schematic. The network is composed of a three-layer convolutional core with a single-point readout predicting neuronal responses, a shifter network accounting for eye movements and a behavioral modulator predicting neuron-specific adaptive gain7,57. Average responses (gray) to test images for two example neurons are plotted with corresponding model predictions (black). e, Performance of the model ensemble, measured as the normalized correlation coefficient between predicted and observed responses to the 100 held-out images (CCnorm)12. Data were pooled over 33,714 neurons from 14 mice (median 0.71, dashed line). Excessively noisy neurons (CCmax < 0.1) were excluded (0.2% of all neurons). Neurons with CCnorm outside [0, 1] were clipped (1.2%) for visualization.
