Deep Lookup Network.

Area of Science:

• Artificial Intelligence
• Computer Science
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are computationally intensive due to operations like multiplication.
• High computational complexity hinders CNN deployment on resource-limited edge devices.
• Multiplication operations in CNNs consume significant energy and increase inference time.

Purpose of the Study:

• To introduce a generic and efficient lookup operation as a basic building block for neural networks.
• To replace computationally expensive multiplication operations with efficient lookup operations.
• To enable end-to-end optimization of the lookup operation for improved network performance.

Main Methods:

• Developed a differentiable lookup table construction for end-to-end training.
• Proposed training strategies to ensure convergence of lookup networks.
• Replaced multiplication operations with lookup operations in CNNs to create lookup networks.
• Applied lookup networks to image classification, image super-resolution, and point cloud classification.

Main Results:

• Lookup networks demonstrate significant improvements in energy consumption and inference speed.
• The proposed lookup operation maintains competitive performance compared to traditional CNNs.
• Achieved state-of-the-art results on various tasks, including classification and regression.
• Validated effectiveness across different data types, such as images and point clouds.

Conclusions:

• Lookup operations offer a viable and efficient alternative to multiplication in CNNs.
• Lookup networks are suitable for deployment on resource-constrained edge devices.
• The approach provides a promising direction for developing more efficient deep learning models.