The paper investigates the transferability of features in deep neural networks, particularly focusing on the transition from general to specific features as the network layers progress. The authors define and quantify the generality and specificity of features at each layer, using a deep convolutional neural network trained on the ImageNet dataset. They find that transferability is negatively affected by two main issues: (1) the specialization of higher-layer neurons to their original task, and (2) optimization difficulties related to splitting networks between co-adapted neurons. The study also reveals that the performance benefits of transferring features decrease as the similarity between the base and target tasks decreases. Surprisingly, initializing a network with transferred features from almost any number of layers can improve generalization performance even after fine-tuning on a new dataset. The findings highlight the importance of understanding the nature and extent of feature transferability for effective transfer learning.The paper investigates the transferability of features in deep neural networks, particularly focusing on the transition from general to specific features as the network layers progress. The authors define and quantify the generality and specificity of features at each layer, using a deep convolutional neural network trained on the ImageNet dataset. They find that transferability is negatively affected by two main issues: (1) the specialization of higher-layer neurons to their original task, and (2) optimization difficulties related to splitting networks between co-adapted neurons. The study also reveals that the performance benefits of transferring features decrease as the similarity between the base and target tasks decreases. Surprisingly, initializing a network with transferred features from almost any number of layers can improve generalization performance even after fine-tuning on a new dataset. The findings highlight the importance of understanding the nature and extent of feature transferability for effective transfer learning.