Recent studies have shown that cortical low-frequency oscillations are often organized as traveling waves. The properties of these waves have been linked to both sensory processing and cognitive functions. In EEG recordings, alpha-band (~10Hz) traveling waves propagate predominantly along the occipital-frontal axis, with forward waves being most prominent during visual processing, while backward waves dominate at rest and during sensory suppression. While a previous study has proposed a functional model to explain their generation and propagation, a biologically plausible implementation is lacking. Here, we present a multi-scale network model with mean-field dynamics that, building on known cortical connectivity, reproduces the dynamics of alpha-band traveling waves observed in EEG recordings. We show that forward and backward waves can arise from two distinct cortical sub-networks that are connected in infragranular layers at each area. At rest, the network generates spontaneous backward waves and switches to a forward state upon sensory stimulation, reproducing the dynamics observed in EEG recordings. We then show that a cortico-thalamic pathway through the pulvinar can bias the dynamics to the forward state and that pulvinar engagement leads to spontaneous forward waves at rest. This is in line with previous studies suggesting a key role for the pulvinar in directing cortical information flow. In summary, our model provides a biologically plausible architecture for modeling the dynamics of macroscale traveling waves. It bridges the gap between scales by connecting laminar activity to scalp-level spatial patterns, providing a biologically grounded and comprehensive view of the spatial propagation of alpha-band traveling waves.

In natural behaviors, multiple neural signals simultaneously drive activation across overlapping brain networks. Due to limitations in the amount of data that can be acquired in common experimental designs, the determination of these interactions is commonly inferred via modeling approaches, which reduce overfitting by finding appropriate regularizing hyperparameters. However, it is unclear whether these hyperparameters can also be related to any aspect of the underlying biological phenomena and help interpret them. We applied a state-of-the-art regularization procedure-automatic locality determination-to interacting neural activations in the mouse posterior cortex associated with movements of the body and eyes. As expected, regularization significantly improved the determination and interpretability of the response interactions. However, regularizing hyperparameters also changed considerably, and seemingly unpredictably, from animal to animal. We found that these variations were not random; rather, they correlated with the variability in visually evoked responses and with the variability in the state of arousal of the animals measured by pupillometry-both pieces of information that were not included in the modeling framework. These observations could be generalized to another commonly used-but potentially less informative-regularization method, ridge regression. Our findings demonstrate that optimal model hyperparameters can be discovery tools that are informative of factors not a priori included in the model's design.

Visual understanding requires comprehending complex visual relations between objects within a scene. Here, we seek to characterize the computational demands for abstract visual reasoning. We do this by systematically assessing the ability of modern deep convolutional neural networks (CNNs) to learn to solve the synthetic visual reasoning test (SVRT) challenge, a collection of 23 visual reasoning problems. Our analysis reveals a novel taxonomy of visual reasoning tasks, which can be primarily explained by both the type of relations (same-different versus spatial-relation judgments) and the number of relations used to compose the underlying rules. Prior cognitive neuroscience work suggests that attention plays a key role in humans' visual reasoning ability. To test this hypothesis, we extended the CNNs with spatial and feature-based attention mechanisms. In a second series of experiments, we evaluated the ability of these attention networks to learn to solve the SVRT challenge and found the resulting architectures to be much more efficient at solving the hardest of these visual reasoning tasks. Most important, the corresponding improvements on individual tasks partially explained our novel taxonomy. Overall, this work provides a granular computational account of visual reasoning and yields testable neuroscience predictions regarding the differential need for feature-based versus spatial attention depending on the type of visual reasoning problem.

Deep neural networks excel at image classification, but their performance is far less robust to input perturbations than human perception. In this work we explore whether this shortcoming may be partly addressed by incorporating brain-inspired recurrent dynamics in deep convolutional networks. We take inspiration from a popular framework in neuroscience: "predictive coding". At each layer of the hierarchical model, generative feedback "predicts" (i.e., reconstructs) the pattern of activity in the previous layer. The reconstruction errors are used to iteratively update the network’s representations across timesteps, and to optimize the network's feedback weights over the natural image dataset--a form of unsupervised training. We show that implementing this strategy into two popular networks, VGG16 and EfficientNetB0, improves their robustness against various corruptions and adversarial attacks. We hypothesize that other feedforward networks could similarly benefit from the proposed framework. To promote research in this direction, we provide an open-sourced PyTorch-based package called \textit{Predify}, which can be used to implement and investigate the impacts of the predictive coding dynamics in any convolutional neural network.