Speakers
Description
Neural network models are primarily characterized by three ingredients: architecture, training objective and training dataset. While research on neural predictivity has largely explored the first two components, dataset choice may offer substantial improvements when modelling the rodent visual cortex, since standard images differ markedly from the low-acuity visual input received by a mouse. Here, we investigated whether an ecologically motivated preprocessing pipeline improved the similarity between representations in mouse visual cortex and convolutional neural networks. We modeled mouse visual input using two transformations: Gaussian blur and pixel-wise Gaussian noise, with parameters tuned to approximate mouse contrast sensitivity. We used the Allen Brain Observatory Visual Coding dataset, which provides neural activity evoked by natural images, and compared these responses to three instances of AlexNet: one trained on ImageNet, one trained on ImageNet processed with the mouse-like pipeline, and an untrained control. Across all visual areas, the model trained on processed images achieved higher neural predictivity than the model trained on unaltered images. However, most of this improvement was recovered by preprocessing images only at inference time: the standard ImageNet-trained model evaluated on processed images performed comparably to the model trained and evaluated with the pipeline. The two Gaussian transformations had distinct effects across network depth: early layers favoured blur alone, consistent with low-pass filtering and reduced visual acuity, whereas deeper layers benefited from noise, possibly by regularizing features shaped by supervised classification. These effects were absent in the untrained model, suggesting that the pipeline acts primarily on meaningful learned representations. Overall, matching image statistics to an animal’s sensory input may provide a simpler and less costly route to improving neural predictivity than redesigning architectures or training objectives. The effects are largely obtained at inference time, with blur and noise making dissociable, depth-dependent contributions.