Recent advances in tau PET imaging offer new insight into the accumulation of neurofibrillary tangles in the brain associated with dementia. Large datasets are now available to create advanced statistical models, such as Deep Learning models. Unfortunately, most of the Deep Networks derived so far from tau PET data were designed to achieve complex tasks, such as classification and regression tasks, that require convoluted analysis to produce interpretable brain maps. In this work, we propose a more straightforward approach to model the presence of tau protein in the brain: we use small interpretable Deep Networks to model the spatial variation across the cortex of probability distributions derived from a large population of 1,386 ADNI tau PET scans. We compare 192 variants of network architectures with standard statistical approaches and select a reliable model to generate brain maps describing how pathological tau proteins accumulate in the cortex.