Splitter: Faster Inference Through Channel Partitioning and Feature Fusion

—This paper presents Splitter, a novel architecture designed to enhance feature extraction and optimize computational efficiency in deep learning models. Splitter employs a unique channel-splitting mechanism that divides input channels into three parallel path; Identity, Activation, and Spatial Mixing to perform distinct operations. By selectively applying spatial mixing via max-pooling or multi-head attention, Splitter balances computational frugality with representational richness. On the ImageNet-1k benchmark, Splitter-S achieves 74.4 % Top-1 accuracy at 9,347 images/s, while Splitter-M and Splitter-L deliver 76.2 % and 78.3 % Top-1 accuracy at 5,893 images/s and 4,719 images/s, respectively. When integrated into a RetinaNet detector on COCO, Splitter-S attains 32.1 % AP (52.4 % AP50, 33.7 % AP75). These results confirm that Splitter matches or surpasses state-of-the-art efficient models while significantly boosting throughput, making it exceptionally well-suited for deployment in resource-limited environments.