Mitigating Resource-Constraints in Computer Vision Systems Using Deep, Sparse CNN Models

Abstract Resource-constrained deployment of object detectors demands models that are simultaneously accurate, robust, and computationally efficient. We investigate whether biologically-inspired activation sparsity—where only a fraction of neurons fire per inference, as in mammalian neocortex—can address all three requirements through a single mechanism. We inject a k-Winner-Take-All (kWTA) activation (retaining only the top-r fraction of channels per spatial location with homeostatic boosting) into the YOLOv11n backbone. Comparing an r=0.5 sparse backbone against an identically configured dense baseline (both trained from scratch, 30 epochs, PASCAL VOC), we find that: (1) with kWTA active at inference, the sparse model substantially outperforms the dense baseline ([email protected]: 0.704 vs. 0.600, +17.3% relative) while simultaneously zeroing 50% of backbone activations, (2) sparse-trained representations are more robust to Gaussian noise, blur, and occlusion (+29.6% relative mean corrupted [email protected], 15/15 conditions favoring the sparse model), (3) an estimated 48.5% of backbone MACs become eliminable under sparsity-aware execution, and (4) both models exhibit negligible degradation under INT8 post-training quantization (<0.001 mAP loss). These results suggest that competitive activation sparsity can simultaneously deliver accuracy improvement and activation-level computational reduction, offering a promising direction for resource-constrained edge deployment.