Awakening Memories from Neural Network Parameters

Neural networks are known to retain memories of their training data, yet how such memories are accessed and used by the model remains poorly understood. In human memory, past experiences can be selectively awakened by visual or auditory cues. Motivated by this analogy, we introduce a new task, termed memory awakening: given a reference input, reconstruct a training memory from the model parameters. To ensure that the awakened memory reflects information stored in the trained model, the task assumes no access to the training data or stored gradient information from training. The expected behavior differs across scenarios: i) a training sample should be able to awaken itself; ii) a partially missing or blurred training sample could awaken a memory that repairs the degraded content; iii) for an unseen input, the awakened memory should be a similar sample, but one from the training set; iv) a sample perturbed by an adversarial attack should awaken a training sample from the target class, even when the two are visually distinct; v) if the learning process is incomplete or disturbed, the awakening results should reflect these defects. To realize this task, we exploit the low-rank structure of parameter updates to construct a memory codebook. Given a reference input, the codebook allows us to selectively awaken a corresponding training memory. Our results consistently support the above behaviors and demonstrate the potential of memory awakening as a new way of understanding neural networks.