A Two-Stage Cascade Model of BOLD Responses in Human Visual Cortex

Abstract

Visual neuroscientists have discovered fundamental properties of neural representation through careful analysis of responses to controlled stimuli. Typically, different properties are studied and modeled separately. To integrate our knowledge, it is necessary to build general models that begin with an input image and predict responses to a wide range of stimuli. In this study, we develop a model that accepts an arbitrary band-pass grayscale image as input and predicts blood oxygenation level dependent (BOLD) responses in early visual cortex as output. The model has a cascade architecture, consisting of two stages of linear and nonlinear operations. The first stage involves well-established computations—local oriented filters and divisive normalization—whereas the second stage involves novel computations—compressive spatial summation (a form of normalization) and a variance-like nonlinearity that generates selectivity for second-order contrast. The parameters of the model, which are estimated from BOLD data, vary systematically across visual field maps: compared to primary visual cortex, extrastriate maps generally have larger receptive field size, stronger levels of normalization, and increased selectivity for second-order contrast. Our results provide insight into how stimuli are encoded and transformed in successive stages of visual processing.

Doubly Sparsifying Network

“…The proposed DSN is implemented with the CUDA ConvNet package [Krizhevsky et al, 2012]. We use a constant learning rate of 0.01, with the momentum parameter fixed at 0.9, and a batch size of 128.…”

“…While many neural networks are trained well with random initializations, it has been discovered that poor initializations can still hamper the effectiveness of first-order methods [Sutskever et al, 2013]. On the other hand, It is much easier to initialize DSN in the right regime.…”

“…Many regularization techniques have been proposed to reduce overfitting in deep learning, such as dropout [Krizhevsky et al, 2012], that set a randomly selected subset of activations to zero within each hidden layer. [Wan et al, 2013] introduced dropconnect for regularizing fully-connected layers, which instead sets a randomly selected subset of weights to zero during training.…”

“…• Baseline II: Baseline I regularized by dropout, with a ratio of 0.5 for each layer [Krizhevsky et al, 2012].…”

“…Unlike common computer vision tasks [Krizhevsky et al, 2012], it is usually infeasible to perform artificial data augmentation or generate "synthetic" data here. Therefore, models that are economic in free parameters are advantageous.…”