Dilated Convolutional Neural Networks for Sequential Manifold-Valued Data

Rudrasis Chakraborty, Xingjian Zhen, Nicholas Vogt, Barbara B. Bendlin, Vikas Singh

TL;DR

This paper introduces a novel Dilated Convolutional Neural Network (DCNN) architecture designed to handle sequential data residing on non-Euclidean manifolds, such as those encountered in neuroimaging. By adapting DCNNs to manifold geometry, the authors achieve state-of-the-art performance in classifying sequential data while significantly reducing computational cost and parameter count, demonstrating its utility in identifying subtle brain connectivity differences related to Alzheimer's disease.

What Problem Does This Paper Address?

Traditional deep learning models, particularly Convolutional Neural Networks (CNNs), excel at processing data with a regular, grid-like structure (like images). However, many real-world datasets, such as those in neuroimaging (e.g., diffusion MRI data representing fiber orientations) or computer vision (e.g., surface normals), do not conform to this Euclidean structure. These "manifold-valued" data points lie on curved spaces where standard operations like convolution are ill-defined. Furthermore, when this data is sequential (e.g., a series of measurements over time or along a path), existing deep learning approaches often rely on Recurrent Neural Networks (RNNs), which can be computationally expensive and difficult to train. This paper addresses the challenge of effectively modeling sequential data that lives on these complex, non-Euclidean manifolds, aiming to leverage the power of deep learning while respecting the underlying geometric structure and improving computational efficiency.

Methods

The core contribution is the development of a Dilated Convolutional Neural Network (DCNN) adapted for manifold-valued data. The authors extend the concept of dilated convolutions, which are known to increase the receptive field of CNNs without increasing parameters, to manifold spaces.

Key methodological innovations include:

  • Manifold-Aware Dilated Convolution: The standard dilated convolution operation is redefined using the Weighted Fréchet Mean (wFM). This operation finds a point on the manifold that minimizes a weighted sum of squared geodesic distances to the input data points. This ensures that the convolution output remains on the manifold.
  • Equivariance to Group Actions: The proposed manifold convolution is shown to be G-equivariant, meaning that if the input data is transformed by a group action (e.g., an isometry), the output is transformed in a corresponding way. This is crucial for preserving geometric properties.
  • Residual Connections for Manifolds: Similar to Euclidean DCNNs, residual connections are incorporated to enable deeper network architectures. These are adapted to the manifold setting by concatenating feature maps and then using wFM to combine them.
  • G-Invariant Last Layer: To ensure the overall network is G-invariant (i.e., the final output doesn't change with group transformations), a specialized last layer is designed. This layer learns group-invariant representations by computing distances to learned manifold means.
  • Weight Normalization, ReLU, and Dropout: The authors argue that due to the properties of wFM (contraction mapping and non-linearity), explicit weight normalization, ReLU activation, and dropout are not strictly necessary, simplifying the architecture.

Datasets

The paper evaluates its method on several datasets:

  • Moving MNIST: A synthetic dataset used to classify moving patterns of digits. This dataset is generated according to a specific algorithm.
  • UCF-11: A real-world dataset for action classification, containing video clips of 11 different actions.
  • Neuroimaging Data: Diffusion-weighted MRI (DWI) data from participants in preclinical Alzheimer's disease studies. This data is used to analyze white matter fiber bundles. The analysis involves two groups: PiB+ (positive for amyloid deposition) vs. PiB- (negative), and APOE+ (carriers of a risk gene) vs. APOE- (non-carriers).

Key Insights and Ideas

  • DCNNs are a viable alternative to RNNs for sequential manifold data: The paper demonstrates that DCNNs can effectively capture temporal dependencies in manifold-valued sequences, offering a computationally more efficient alternative to RNNs.
  • Geometric awareness is key: Explicitly incorporating the manifold structure (via wFM) into the convolution operation leads to better performance and more meaningful results, especially in scientific applications.
  • Simplification through geometry: The inherent properties of the manifold-based wFM operation reduce the need for standard regularization techniques like weight normalization and dropout, simplifying the model.
  • Enabling scientific discovery: The method's ability to identify subtle differences in brain connectivity in cognitively healthy individuals at risk for Alzheimer's disease highlights its potential for early disease detection and understanding complex biological processes.

Key Results

  • Improved Efficiency: The manifold DCNN achieved comparable or better classification accuracy than state-of-the-art methods (like SPD-SRU, LSTM, TT-GRU) on Moving MNIST and UCF-11 datasets, while using significantly fewer parameters and running much faster (e.g., ~2.5x faster than SPD-SRU on UCF-11, and 5x faster in permutation testing for neuroimaging).
  • Neuroimaging Findings: In the Alzheimer's disease study, the model identified significant differences in several white matter fiber bundles between groups (PiB+ vs. PiB-, APOE+ vs. APOE-). Notably, it found differences in bundles known to be affected by AD pathology, such as the superior longitudinal fasciculus and cingulum bundle, and was more sensitive than a baseline SPD-SRU model.
  • Robustness: The model demonstrated robustness to variations in orientation differences in the Moving MNIST dataset, maintaining high accuracy.

Limitations and Future Work

  • Uncorrected p-values: The statistical analysis for neuroimaging data used uncorrected p-values, which is a limitation due to multiple testing. The authors acknowledge this and suggest that multiple testing correction would be a natural next step.
  • Generalization to other manifolds: While the paper focuses on SPD matrices and spheres (for ODFs), future work could explore extending the framework to a wider range of Riemannian manifolds.
  • Further exploration of neuroimaging applications: The authors suggest that their method can facilitate further scientific discovery in neuroimaging and potentially aid in early disease detection.

Summary updated

BibTeX
@inproceedings{chakraborty2019dilated,
  author = {Rudrasis Chakraborty and Xingjian Zhen and Nicholas Vogt and Barbara B. Bendlin and Vikas Singh},
  title = {{Dilated Convolutional Neural Networks for Sequential Manifold-Valued Data}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2019},
  doi = {10.1109/iccv.2019.01072},
}
Figure