Exploring the efficacy of transfer learning in mining image-based software artifacts

Abstract

Despite the recent successes of deep architectures, such as convolutional neural networks, on software engineering data, the lack of sufficiently large training sets for some applications continues to be a substantial hurdle. This requirement has led researchers to label tens of thousands [1] and even millions of images [2] by hand. Recent work has shown that this precludes the use of many off-the-shelf convolutional neural network architectures, requiring empirical software engineering researchers to rely on custom (more compact) architectures [3]. Another possible solution, however, is to leverage

Transfer Learning in Breast Cancer Diagnoses via Ultrasound Imaging

“…To prevent changing the transferred parameters too early, it is customary to start with frozen parameters [40][41][42][43][44], train only randomly initialized layers until they converge, and then unfreeze all parameters and fine-tune ( Figure 1) the entire network. Transfer learning is particularly useful when there is a limited amount of data for one task and a large volume of data for another similar task, or when there exists a model that has already been trained on such data [45]. However, even if there is sufficient data for training a model from scratch and the tasks are not related, initializing the parameters using a pre-trained model is still better than random initialization [46].…”

“…Particularly, all convolution layers of the well-trained CNN model are fixed, whereas the fully connected layers are cleared up [31][32][33][34][35][36][37][38][39]. The convolution layers are used as a frozen feature extractor to match with a new task, such as a breast cancer classification task [41][42][43][44][45]. The extracted features are then supplied to a classifier that can form fully connected layers [45].…”

“…The convolution layers are used as a frozen feature extractor to match with a new task, such as a breast cancer classification task [41][42][43][44][45]. The extracted features are then supplied to a classifier that can form fully connected layers [45]. Lastly, the new classifier is only trained throughout the training process instead of the entire network [51][52][53].…”

“…This is implemented by initializing the weights of the convolution layers of a new model with the pre-trained weights of the already well-trained CNN model, and initializing the classifier layers with arbitrary random weights. In fine-tuning, the entire network is trained during the training process [41][42][43][44][45].…”

Automatic Classification of UML Class Diagrams Using Deep Learning Technique: Convolutional Neural Network

“…Best et al [31] explored the applicability of transfer learning utilizing models that are pre-trained on non software engineering data to the problem of classifying UML diagrams. Their results show training reacts positively to transfer learning as related to sample size, even though the pre-trained model was not exposed to training instances from the software domain.…”