Update README.md

README.md

@@ -34,8 +34,9 @@ tags:
   - Nesterov: True  
 - **Scheduler**: CosineAnnealingLR (T_max: 200)  
 - **Training Epochs**: 62  
-- **Batch Size**: 128  
-- **Hardware**: Single GPU (NVIDIA GeForce RTX 3090)  
+- **Batch Size**: 64  
+- **Hardware**: Single GPU (NVIDIA GeForce RTX 3090)
+- **Number of Retrain**: 1  
 
 ### Algorithm
 The **CF-k** algorithm was used for inexact unlearning. This method systematically removes the influence of a specific class from the model while retaining the ability to classify the remaining classes. Each resulting model (`cifar10_resnet18_CF-k_X.pth`) corresponds to a scenario where a single class (`X`) has been unlearned. The CF-k algorithm provides an efficient framework for evaluating the robustness and adaptability of models under inexact unlearning constraints.
@@ -46,53 +47,18 @@ For more details on the CF-k algorithm, refer to the [GitHub repository](https:/
 
 ## Results
 
-| Model                          | Excluded Class | Forget class acc(loss) | Retain class acc(loss) |
-|--------------------------------|----------------|-------------------------|-------------------------|
-| cifar10_resnet18_CF-k_0.pth    | Airplane       | 0.0 (4.993)            | 81.22 (0.578)          |
-| cifar10_resnet18_CF-k_1.pth    | Automobile     | 0.0 (4.546)            | 95.41 (0.174)          |
-| cifar10_resnet18_CF-k_2.pth    | Bird           | 0.0 (4.819)            | 95.93 (0.154)          |
-| cifar10_resnet18_CF-k_3.pth    | Cat            | 0.0 (5.098)            | 96.54 (0.122)          |
-| cifar10_resnet18_CF-k_4.pth    | Deer           | 0.0 (4.526)            | 95.49 (0.165)          |
-| cifar10_resnet18_CF-k_5.pth    | Dog            | 0.0 (4.843)            | 96.28 (0.134)          |
-| cifar10_resnet18_CF-k_6.pth    | Frog           | 0.0 (4.348)            | 95.31 (0.176)          |
-| cifar10_resnet18_CF-k_7.pth    | Horse          | 0.0 (4.440)            | 95.26 (0.175)          |
-| cifar10_resnet18_CF-k_8.pth    | Ship           | 0.0 (4.453)            | 95.44 (0.171)          |
-| cifar10_resnet18_CF-k_9.pth    | Truck          | 0.0 (4.657)            | 95.59 (0.167)          |
+| Model                          | Forget Class | Forget class acc(loss) | Retain class acc(loss) |
+|--------------------------------|--------------|-------------------------|-------------------------|
+| cifar10_resnet18_CF-k_0.pth    | Airplane     | 0.0 (4.659)             | 95.49 (0.168)          |
+| cifar10_resnet18_CF-k_1.pth    | Automobile   | 0.0 (4.571)             | 95.34 (0.181)          |
+| cifar10_resnet18_CF-k_2.pth    | Bird         | 0.0 (4.879)             | 95.89 (0.158)          |
+| cifar10_resnet18_CF-k_3.pth    | Cat          | 0.0 (5.165)             | 96.56 (0.127)          |
+| cifar10_resnet18_CF-k_4.pth    | Deer         | 0.0 (4.562)             | 95.52 (0.170)          |
+| cifar10_resnet18_CF-k_5.pth    | Dog          | 0.0 (4.862)             | 96.30 (0.137)          |
+| cifar10_resnet18_CF-k_6.pth    | Frog         | 0.0 (4.458)             | 95.37 (0.185)          |
+| cifar10_resnet18_CF-k_7.pth    | Horse        | 0.0 (4.514)             | 95.23 (0.179)          |
+| cifar10_resnet18_CF-k_8.pth    | Ship         | 0.0 (4.577)             | 95.38 (0.178)          |
+| cifar10_resnet18_CF-k_9.pth    | Truck        | 0.0 (4.644)             | 95.53 (0.174)          |
 
 
----
-
-### Notes
-
-1. **Forget Class Accuracy and Loss**:
-   - Across all excluded classes, the forget class accuracy is consistently `0.0`, demonstrating the effectiveness of the **CF-k** method in completely excluding the target classes.
-   - The forget class loss varies slightly, ranging from `4.348` ("Frog") to `5.098` ("Cat"), suggesting that some classes might be slightly more challenging to suppress completely in terms of loss.
-
-2. **Retain Class Accuracy and Loss**:
-   - The retain class accuracy is consistently high across all excluded classes, ranging from `81.22%` ("Airplane") to `96.54%` ("Cat"). This indicates that the method effectively preserves performance on the remaining classes.
-   - Retain class loss is minimal, with the lowest being `0.122` for "Cat" and the highest being `0.578` for "Airplane." This suggests that the model maintains stable performance with minimal degradation on the retained classes.
-
-3. **Class-Specific Observations**:
-   - "Cat" shows the highest retain class accuracy (96.54%) and the lowest retain class loss (0.122), making it the least affected by the exclusion of other classes.
-   - "Airplane" exhibits the lowest retain class accuracy (81.22%) and the highest retain class loss (0.578), indicating a potential trade-off in preserving performance for this class.
-   - Variations in retain class accuracy and forget class loss across different excluded classes highlight the potential influence of class-specific features on model performance.
-
----
-
-### Conclusion
-
-The results illustrate that the **CF-k method** is highly effective in achieving class-specific exclusion while maintaining strong performance on the retained classes. However, minor variations in performance across classes reveal opportunities for further refinement:
-
-- **Strengths**:
-  - The forget class accuracy remains at `0.0` for all excluded classes, achieving complete suppression of the target classes.
-  - Retain class accuracy is high across the board, with most classes exceeding `95%`, showing the robustness of the method in retaining knowledge.
-
-- **Weaknesses**:
-  - "Airplane" has noticeably lower retain class accuracy (81.22%) and higher retain class loss (0.578), indicating that certain classes may be more challenging to balance during the exclusion process.
-  - Slight variations in forget class loss suggest that the suppression process may not be uniformly effective across all classes.
-
-- **Future Work**:
-  - Investigate why certain classes, such as "Airplane," are more impacted in terms of retain class accuracy and loss. Class-specific characteristics or relationships with other classes might influence this outcome.
-  - Explore adaptive mechanisms to optimize the trade-off between exclusion and retention for more balanced performance across all classes.
-  - Conduct additional experiments to determine if similar patterns emerge in other datasets or architectures, which could validate the generalizability of the method.