Learning without Forgetting (LwF) is a method for training Convolutional Neural Networks (CNNs) to learn new tasks while preserving performance on existing tasks without access to the original training data. The method uses only new task data to train the network, ensuring that the original capabilities are maintained. LwF outperforms feature extraction and fine-tuning in new task performance and performs similarly to multitask learning that uses original task data. It also acts as a regularizer to improve new task performance. The method involves training a CNN with shared parameters $ \theta_{s} $, task-specific parameters for existing tasks $ \theta_{o} $, and new task-specific parameters $ \theta_{n} $. The goal is to optimize both new task accuracy and preservation of original task outputs. LwF uses a knowledge distillation loss to encourage outputs on the original tasks to remain similar to the original network. This approach avoids catastrophic forgetting and allows for efficient training without requiring access to original task data. LwF is compared to other methods such as feature extraction, fine-tuning, and joint training. It is faster to train than joint training and requires less computational resources. It also simplifies deployment since training data does not need to be retained after learning a new task. LwF outperforms fine-tuning on new tasks and performs well on both new and old tasks. It is particularly effective when the new task is similar to the original task. Experiments show that LwF performs well on various image classification tasks, including VOC, CUB, and Scenes. It outperforms fine-tuning and feature extraction in most scenarios. The method is also effective in tracking applications, where it can be used to maintain performance on existing tasks while learning new ones. LwF is a promising approach for continual learning and has implications for applications where access to original task data is not available.Learning without Forgetting (LwF) is a method for training Convolutional Neural Networks (CNNs) to learn new tasks while preserving performance on existing tasks without access to the original training data. The method uses only new task data to train the network, ensuring that the original capabilities are maintained. LwF outperforms feature extraction and fine-tuning in new task performance and performs similarly to multitask learning that uses original task data. It also acts as a regularizer to improve new task performance. The method involves training a CNN with shared parameters $ \theta_{s} $, task-specific parameters for existing tasks $ \theta_{o} $, and new task-specific parameters $ \theta_{n} $. The goal is to optimize both new task accuracy and preservation of original task outputs. LwF uses a knowledge distillation loss to encourage outputs on the original tasks to remain similar to the original network. This approach avoids catastrophic forgetting and allows for efficient training without requiring access to original task data. LwF is compared to other methods such as feature extraction, fine-tuning, and joint training. It is faster to train than joint training and requires less computational resources. It also simplifies deployment since training data does not need to be retained after learning a new task. LwF outperforms fine-tuning on new tasks and performs well on both new and old tasks. It is particularly effective when the new task is similar to the original task. Experiments show that LwF performs well on various image classification tasks, including VOC, CUB, and Scenes. It outperforms fine-tuning and feature extraction in most scenarios. The method is also effective in tracking applications, where it can be used to maintain performance on existing tasks while learning new ones. LwF is a promising approach for continual learning and has implications for applications where access to original task data is not available.