Yurim0507/Selective_Synapse_Dampening

Evaluation Report

Testing Data

Dataset: CIFAR-10 Test Set
Metrics: Forget class accuracy(loss), Retain class accuracy(loss)

---

Training Details

Training Procedure

• Base Model: ResNet18
• Dataset: CIFAR-10
• Excluded Class: Varies by model
• Loss Function: Negative Log-Likelihood Loss
• Optimizer: SGD with:
  • Learning rate: 0.01
  • Momentum: 0.9
  • Weight decay: 5e-4
  • Nesterov: True
• Training Epochs: 20
• Batch Size: 256
• Hardware: Single GPU (NVIDIA GeForce RTX 3090)

Selective Synapse Dampening Specifics

• Lambda: 1.0
• Alpha: 10.0

Algorithm

The SSD (Selective Synapse Dampening) algorithm was used for inexact unlearning. This method selectively reduces the impact of a specific class on the model while preserving the performance on the remaining classes.

Each resulting model (cifar10_resnet18_SSD_X.pth) corresponds to a scenario where a single class (X) has been unlearned. SSD efficiently removes class-specific knowledge while maintaining robustness and generalizability.

For more details on the SSD algorithm, refer to the GitHub repository.

---

Results

Model	Forget Class	Forget class acc(loss)	Retain class acc(loss)
cifar10_resnet18_SSD_0.pth	Airplane	0.0 (8.102)	83.38 (0.527)
cifar10_resnet18_SSD_1.pth	Automobile	0.0 (6.550)	94.62 (0.189)
cifar10_resnet18_SSD_2.pth	Bird	0.0 (9.854)	90.06 (0.328)
cifar10_resnet18_SSD_3.pth	Cat	0.0 (8.428)	90.00 (0.317)
cifar10_resnet18_SSD_4.pth	Deer	0.0 (5.885)	95.26 (0.161)
cifar10_resnet18_SSD_5.pth	Dog	0.0 (6.917)	12.53 (2.799)
cifar10_resnet18_SSD_6.pth	Frog	0.0 (5.532)	95.29 (0.156)
cifar10_resnet18_SSD_7.pth	Horse	0.0 (7.328)	17.71 (3.478)
cifar10_resnet18_SSD_8.pth	Ship	0.0 (3.783)	95.41 (0.158)
cifar10_resnet18_SSD_9.pth	Truck	0.0 (5.864)	94.29 (0.198)