Learnable Heterogeneous Convolution: Learning both topology and strength.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Deep Learning

Background:

• Traditional convolution techniques in artificial neural networks (ANNs) face substantial computational challenges.
• Biological neural networks (BNNs) exhibit remarkable efficiency and power, offering a model for optimization.

Purpose of the Study:

• To develop a novel convolution method inspired by BNNs to address the computational complexity of ANNs.
• To unify existing handcrafted convolution techniques through a data-driven approach.

Main Methods:

• Introduced Learnable Heterogeneous Convolution, enabling joint learning of kernel shape and weights.
• Leveraged biological plasticity of dendritic topology and synaptic strength as inspiration.
• Achieved structural sparse weights for accelerated computation on parallel devices.

Main Results:

• Reduced computation by approximately 5× on CIFAR10 and 2× on ImageNet for models like VGG16/19 and ResNet34/50 without performance degradation.
• Achieved weight compression of 10× on CIFAR10 and 4× on ImageNet.
• Improved accuracy by up to 1.0% on CIFAR10 and 0.5% on ImageNet with enhanced efficiency.

Conclusions:

• Learnable Heterogeneous Convolution offers a powerful and efficient alternative to existing convolution methods in ANNs.
• The method demonstrates significant potential for accelerating deep learning models while maintaining or improving performance.