SepNet: A neural network for directionally correlated data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multi-dimensional tensor data are prevalent in fields like signal processing and remote sensing.
• Directionally correlated data exhibit stronger intra-dimensional than inter-dimensional correlations.
• Existing convolutional neural networks (CNNs) are inefficient for such data due to excessive neuron connections.

Purpose of the Study:

• To introduce SepNet, a novel neural network architecture tailored for directionally correlated tensor data.
• To enhance the efficiency and maintain the accuracy of deep learning models on specific data structures.
• To demonstrate the benefits of data-specific neural network design.

Main Methods:

• SepNet employs directional operators to process each dimension independently.
• It utilizes a linear operator along the depth to integrate directional features into higher-level representations.
• The architecture allows flexible construction with minimal output shape constraints.

Main Results:

• SepNet achieved up to 100-fold improvement in network efficiency compared to standard CNNs.
• The proposed model maintained high accuracy, comparable to state-of-the-art CNNs.
• Experiments were conducted on two representative directionally correlated datasets.

Conclusions:

• SepNet offers a highly efficient and accurate solution for processing directionally correlated tensor data.
• The findings highlight the potential of architecting neural networks based on specific data characteristics.
• This approach paves the way for more specialized and performant deep learning models.