The Architecture of Neural Encoding: Comparing Deep Learning Models of Visual Cortex

Abstract

Previous work in neural encoding has focused on exploring the potential of deep learning architecture in modeling sensory perception, with several distinct forms of deep neural networks rising as particular computational representations of underlying cortical processes. These encoding strategies, however, remain categorically ambiguous due to the difficulty in mapping nonlinear transformative layers in neural networks to the neurobiological equipment responsible for analogous calculations. This is a problem fundamental to neural networks, however, due to the nature of their interconnected and multilayered architecture, which makes the interpretation of any discrete layer a difficult analytical task. Moreover, this obstacle is one that must be overcome in order to improve the performance of neural activity models in general, and thus this dilemma lies at the foundation of much research currently progressing in the field of computational neuroscience. Two specific and organizationally distinct deep learning approaches to modeling visual neural data have recently demonstrated not only predictive success, but performance improvements over earlier methods for modeling neuron responses in visual sensory areas. Here, we seek to create, train, and evaluate two separate machine learning-based model architectures using recording data from two, hierarchically-linked cell classes in the visual processing hierarchy. Results indicate that while the addition of individual convolutional filter nodes may not directly improve model accuracy, such steps are nonetheless important for enriching a network’s feature space and are critical to enhancing prediction accuracy for computationally complicated neural substrates.