In this work, we identify the heterogeneity of regions of white matter across individuals and the complexity of different regions of white matter using diffusion MRI data. We analyze the heterogeneity of a region across individuals by computing the pairwise difference between voxels. Using autoencoders, we analyze the inherent dimensionality of regions of white matter structure and find that some regions are easier to compress than others. The intrinsic complexity of a region can be determined by how effectively it can be predicted given its neighborhood. We pose this as the problem of inpainting a three dimensional region of the brain given its context with deep generative modelling using Generative Adversarial Networks (GANs) conditioned on the neighborhood. The discriminator is trained on differentiating between real and generated patches while the generator is simultaneously trained to adversarially fool the discriminator and minimize the voxel-wise reconstruction loss between the actual and generated patches. We study the comparative performance of a well-trained GAN across regions of the brain in multiple subjects and identify particularly challenging or complex regions of brain wiring.