Learnable pairwise filter connections boost CNN accuracy

Traditional convolutional neural networks stack convolutional blocks separated by pointwise activation functions (e.g., ReLU). While effective, this pattern limits the network's ability to model complex interactions between filters. The authors introduce a novel alternative: adding learnable connections between pairs of filters. Instead of using a fixed operation like multiplication or the minimum (which act as logical AND gates), their method parameterizes each connection, allowing the network to learn the optimal function per layer and per filter pair.

In experiments on standard vision benchmarks, the proposed approach consistently outperforms baseline CNNs that rely solely on pointwise nonlinearities. The learnable connections enable the network to implement diverse operations (e.g., weighted sums, gating) tailored to the specific task. This flexibility leads to higher accuracy with minimal overhead. The work, accepted at IJCNN 2023, opens a new direction for designing CNN architectures beyond conventional activation functions, offering a simple yet powerful mechanism to boost performance.