CQNN

Standard convolutional neural networks rely on linear transformations followed by an activation function (like ReLU). CQNN rethinks this fundamental building block by introducing Quadratic Neurons, which inherently capture second-order interactions between input features.

Why Quadratic Neurons?

In many computer vision tasks, the decision boundaries are naturally non-linear. By using a quadratic form $y = x^T W x$ as the basic operation, CQNNs can:

• Increase Model Capacity: Capture complex patterns with fewer total layers.
• Improve Decision Boundaries: Achieve better separation in high-dimensional feature spaces compared to standard linear-based layers.
• Robust Feature Extraction: Enhance performance in texture analysis and intricate pattern recognition tasks.

Key Contributions

• Non-Linear Primitives: Integration of quadratic operations directly into the convolutional framework.
• Experimental Validation: Demonstrated superior performance on standard benchmarks for image classification and pattern recognition.
• Efficiency: Achieving high accuracy with a more compact representation of non-linearities.

Resources