1: \begin{abstract}

2: Efforts are underway to study ways via which the power

3: of deep neural networks can be extended to non-standard data types such as structured data (e.g., 

4: graphs) or manifold-valued data (e.g., unit vectors or special matrices).

5: %New results show that

6: Often, sizable empirical improvements are possible when the geometry of such

7: data spaces are incorporated into the design of the model, architecture, and the algorithms.

8: Motivated by neuroimaging applications,

9: we study formulations where the data are {\em sequential manifold-valued

10: measurements}. This case is common in brain imaging, where the samples correspond to

11: symmetric positive definite matrices or orientation distribution functions.

12: Instead of a recurrent model which poses computational/technical issues,

13: and inspired by recent results showing the viability of dilated convolutional models for

14: sequence prediction, we develop a dilated convolutional neural network architecture

15: for this task. On the technical side, we show how the modules needed in our network

16: can be derived while explicitly taking the Riemannian manifold structure into account.

17: We show how the operations needed can leverage known results

18: for calculating the weighted Fr\'{e}chet Mean (wFM).

19: Finally, we present scientific results for group difference analysis in Alzheimer's disease (AD)

20: where the groups are derived using AD pathology load: here the model finds

21: several brain fiber bundles that are related to

22: AD even when the subjects are all still cognitively healthy. 

23: \end{abstract}

24: