Memory-Modulated CNNs: Simulating Cognitive Interference via Decaying Feature Recall

In this experiment, we design a memory-modulated Convolutional Neural Network (CNN) architecture where the model's standard inference behavior is altered by previously stored "memories." These memories are generated by capturing the intermediate outputs (feature maps) from the first and second convolutional layers when an image is passed through the model. Once stored, these memory tensors are subject to linear degradation each time the model performs inference, simulating the natural fading of memory over time. Additionally, to emulate fuzzy recall observed in human cognition, one random channel of each memory tensor is shuffled during each inference pass. Experimental results demonstrate that the injection of undegraded memories into the model significantly worsens inference performance when compared to the baseline model. However, as the stored memories degrade over repeated inference cycles, the model's performance approaches that of the original baseline, indicating that degradation mitigates the disruptive influence of memory injection.