Tiny Data Is Sufficient: A Generalizable CNN Architecture for Temporal Domain Long Sequence Identification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep learning (DL) models excel at sequence processing but require extensive data and parameters.
• Conventional convolutions in deep networks can limit feature representation for long temporal sequences.
• Existing models often struggle with feature processing efficacy in long sequence analysis.

Purpose of the Study:

• Introduce a novel generalizable convolutional neural network (GeCNN) architecture.
• Address challenges in long temporal sequence identification with limited data.
• Improve feature representation and accuracy in deep learning models.

Main Methods:

• Developed a GeCNN framework with generic CNN, selective CNN, and multiple pooling layers.
• Incorporated customizable hyper-convolutional operations via non-linear convolvers.
• Utilized selective CNN with homogeneous striding principle and partial homogeneous striding theorem to reduce data dependency.
• Combined eight distinct pooling operations to minimize statistical information loss.

Main Results:

• GeCNN demonstrated superior performance with shallow networks and small datasets compared to deep networks.
• Achieved higher accuracy than ResNet and self-attention models on the GTZAN dataset using significantly less training data.
• Outperformed other models on the PLAID dataset with minimal data requirements.

Conclusions:

• The proposed GeCNN architecture offers a powerful solution for temporal domain long sequence identification.
• GeCNN effectively enhances feature representation and accuracy while reducing the need for large training datasets.
• This approach presents a significant advancement for efficient and accurate deep learning in sequence analysis.