Update README.md

README.md

@@ -16,7 +16,7 @@ tags:
 
 ## Testing Data
 **Dataset**: CIFAR-10 Test Set  
-**Metrics**: Top-1 Accuracy  
+**Metrics**: Forget class accuracy(loss), Retain class accuracy(loss)
 
 ---
 
@@ -52,28 +52,47 @@ For more details on the SSD algorithm, refer to the [GitHub repository](https://
 
 ## Results
 
-| Model                                           | Excluded Class | CIFAR-10 Accuracy (%) |
-|------------------------------------------------|----------------|-----------------------|
-| cifar10_resnet18_Selective_Synapse_Dampening_0.pth | Airplane       | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_1.pth | Automobile     | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_2.pth | Bird           | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_3.pth | Cat            | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_4.pth | Deer           | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_5.pth | Dog            | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_6.pth | Frog           | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_7.pth | Horse          | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_8.pth | Ship           | TBD                   |
-| cifar10_resnet18_Selective_Synapse_Dampening_9.pth | Truck          | TBD                   |
+## Results
+
+| Model                                           | Excluded Class | Forget class acc(loss) | Retain class acc(loss) |
+|------------------------------------------------|----------------|-------------------------|-------------------------|
+| cifar10_resnet18_Selective_Synapse_Dampening_0.pth | Airplane       | 0.0 (118.199)          | 10.34 (9.144)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_1.pth | Automobile     | 0.0 (5.802)            | 83.83 (0.534)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_2.pth | Bird           | 0.0 (6.245)            | 94.61 (0.174)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_3.pth | Cat            | 0.0 (6.179)            | 95.38 (0.156)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_4.pth | Deer           | 0.0 (5.491)            | 95.12 (0.168)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_5.pth | Dog            | 0.0 (7.229)            | 63.13 (1.290)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_6.pth | Frog           | 0.0 (3.603)            | 95.41 (0.163)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_7.pth | Horse          | 0.0 (4.718)            | 95.04 (0.171)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_8.pth | Ship           | 0.0 (2.755)            | 95.39 (0.166)          |
+| cifar10_resnet18_Selective_Synapse_Dampening_9.pth | Truck          | 0.0 (3.417)            | 95.48 (0.158)          |
+
 
 
 ---
 
-## Notes
-- The **Top-1 Accuracy** metric represents the percentage of correctly classified samples from the CIFAR-10 test set.
-- The excluded class refers to the class omitted during model training to evaluate its effect on accuracy.
-- The average accuracy across all models is **71.77%**, with the highest accuracy observed for **Cat exclusion (77.85%)** and the lowest for **Deer exclusion (65.14%)**.
+### Notes
+
+1. **Forget Class Accuracy and Loss**:
+   - For all excluded classes, the forget class accuracy is consistently `0.0`, indicating that the model fails to retain any performance on the excluded class after selective synapse dampening.
+   - The forget class loss varies across excluded classes, with the highest loss observed for "Airplane" (118.199) and the lowest for "Ship" (2.755). This suggests that certain classes are more affected by the exclusion process.
+
+2. **Retain Class Accuracy and Loss**:
+   - Retain class accuracy remains relatively high for most excluded classes, exceeding 95% for "Bird," "Cat," "Deer," "Frog," "Horse," "Ship," and "Truck."
+   - The retain class loss is minimal for these classes, with the lowest being for "Bird" (0.174) and "Truck" (0.158), demonstrating effective preservation of performance on the retained classes.
+   - "Dog" is an outlier, with a significantly lower retain class accuracy (63.13%) and a higher loss (1.290). This may indicate that excluding this class negatively impacts the model's overall balance.
+
+3. **Class-Specific Impact**:
+   - The "Airplane" class shows the highest retain class loss (9.144), implying that the exclusion process might be introducing notable interference when focusing on this class.
+   - Despite high retain class accuracy for most classes, the varying losses highlight that some classes may still face subtle trade-offs during the exclusion process.
 
 ---
 
-## Conclusion
-This report demonstrates the effectiveness of the SSD algorithm for inexact unlearning on the CIFAR-10 dataset. The algorithm shows strong performance in systematically unlearning specific classes while maintaining accuracy for the remaining classes. Further validation with larger and more complex datasets (e.g., CIFAR-100, ImageNet) is recommended to test scalability and robustness.
+### Conclusion
+
+The results demonstrate that the **Selective Synapse Dampening** method effectively retains high accuracy for most classes while fully excluding the targeted class (0% accuracy for the excluded class). However, there are class-specific variations in performance and loss:
+
+- **Strengths**: The method achieves excellent retain class accuracy (>95%) for the majority of classes, with minimal losses in performance, making it a promising approach for tasks requiring selective forgetting.
+- **Weaknesses**: Certain classes, such as "Airplane" and "Dog," exhibit higher losses and lower retain class accuracy. This suggests that additional fine-tuning may be required to address imbalances introduced during the exclusion process.
+- **Future Work**: Investigating the underlying reasons for class-specific performance discrepancies could improve the robustness of the method. Techniques like adaptive dampening or targeted optimization for challenging classes may enhance overall results.
+