Deep Convolutional Tables: Deep Learning Without Convolutions.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional layers are a computational bottleneck in deep learning, limiting applications on resource-constrained devices like IoT and CPUs.
• Existing deep learning models often rely on dot-product neurons, which can be computationally intensive.

Purpose of the Study:

• To propose a novel deep network formulation using convolutional tables (CTs) for accelerated CPU-based inference.
• To demonstrate that CTs can overcome the limitations of traditional convolutional layers in terms of computational complexity and efficiency.

Main Methods:

• Developed convolutional tables (CTs) that utilize a hierarchy of voting tables instead of dot-product neurons.
• Each CT performs a fern operation, encoding local image environments into binary indices for table lookups.
• Derived a soft relaxation and gradient-based training approach for the CT hierarchy.

Main Results:

• CT computational complexity is independent of patch size and scales gracefully with channels, outperforming standard convolutional layers.
• Deep CT networks exhibit a universal approximation property and comparable accuracy to CNNs.
• CTs provide a superior error-speed tradeoff in low-compute regimes compared to efficient CNNs.

Conclusions:

• Convolutional tables offer a promising alternative to traditional convolutional layers for efficient deep learning inference on CPUs.
• CT networks achieve competitive accuracy while significantly improving computational efficiency, particularly for edge devices.
• The proposed CT formulation enables effective deep learning in environments with limited computational resources.