Project Summary/Abstract Recent technical advances have enabled to record the activity of hundreds or thousands of neurons simulta- neously. To extract insight from this deluge of data, computational neuroscientists have turned to data-driven modeling approaches. In this paradigm, an artificial recurrent neural network (RNN) is first fit to imitate recorded activity, and is then dissected to reveal putative mechanisms. A prominent goal of this approach is to uncover the dynamics of computation in the trained RNN, including attractor structure that defines how the network could accumulate a signal or stably store a memory. However, a firm theoretical understanding of these RNN models is lacking. In particular, are they guaranteed to uncover true mechanisms, or can they find spurious structure? How does the RNN architecture chosen affect how they learn to imitate observed dynamics? The proposed research aims to resolve these foundational gaps in our understanding of a widely used approach to extracting insight from high-dimensional neural data. First, it aims to establish the fundamental domain of validity of recovering attractor structure from data by developing benchmarks that can be applied to any data-driven method. This represents a shift in how data-driven models are evaluated, from focusing on their ability to explain variance in test data to instead demanding that they robustly uncover underlying mechanisms. Second, it aims to advance our ba- sic understanding of how RNNs learn to mimic observed dynamics. Using the powerful toolkit of modern deep learning theory, I aim to build a more complete theory of how network architecture and training procedure interact to bias how an RNN imitates real neural dynamics. In total, this research will elucidate the limitations of one of the most popular approaches for extracting understanding from large-scale neural data, and advance our basic understanding of how recurrent computations are learned.