PROJECT SUMMARY In population genetics, deep neural networks have largely replaced traditional approaches such as approximate Bayesian computation (ABC) and other summary statistic-based methods. However, there has been a lack of exploration and comparison of different architectures and input representation of genetic data. Additionally, interpreting trained neural networks and their predictions remains largely unexplored in this field. In this project we will develop new inference frameworks based on transformers, and compare these approaches with existing methodology. We will focus on the challenging task of joint inference, where a locus-specific forces such as natural selection or per site, per haplotype admixture are confounded by demographic history. We will create interpretability frameworks that allow researchers to understand what deep learning methods are learning at multiple scales. Local interpretability will explain network output on the genetic region level, while global interpretability focuses on individual network components such as convolutional filters or fully connected layers. Our interpretability methods will allow us to generate not just predictions and parameter estimates, but higher-level scientific insight for diverse human populations. Taken together, outcomes from our results will guide researchers toward the best networks for their inference tasks, quantify evolutionary processes in admixed populations, and illuminate the reasoning behind network predictions.