Fig. 2
From: Modifier guided resilient CNN inference enables fault-tolerant edge collaboration for IoT
Output aggregation strategies for fusing device predictions. (a) Average Pooling: probabilities from all devices are averaged class-wise, yielding a smoothed consensus output (e.g., the final probabilities for Class 1, Class 2, and Class 3 are 0.43, 0.27, 0.30 in this example). (b) Max Pooling: the highest probability from any device is taken for each class (resulting in final probabilities 0.70, 0.50, 0.40 here). Max pooling prioritizes any strong confidence from individual devices, whereas average pooling balances contributions from all devices.
