Tied Block Convolution: Leaner and Better CNNs with Shared Thinner Filters

Convolution is the main building block of a convolutional neural network (CNN). We observe that an optimized CNN often has highly correlated filters as number channels increases with depth, reducing expressive power feature representations. propose Tied Block (TBC) shares same thinner filter over equal blocks and produces multiple responses single filter. The concept TBC can also be extended to group convolution fully connected layers, applied various backbone networks attention modules. Our extensive experimentation on classification, detection, instance segmentation, demonstrates consistently leaner significantly better than standard convolution. On attention, 64 times fewer parameters, our TiedSE performs par SE. detection effectively handle overlapping instances, whereas CNNs fail accurately aggregate information in presence occlusion result redundant partial object proposals. By sharing across channels, reduces correlation delivers sizable gain 6% average precision for MS-COCO when ratio 80%.