Data Selection via Semantic Similarity for Resource-Efficient Transfer Learning of Image Classifiers

The generation of data such as text, audio and video has rapidly increased over the past 20 years through technological advancements such as IoT (internet of things) and social networks. This data contains information that can help make educated decisions by providing helpful insights. For example, social media data can provide ads tailored for users, leading to more effective advertising. On the other hand, data generated by sensors in IoT devices can tell manufacturers how to enhance the performance of their products. Hence, there is much interest in extracting the value present in data. However, said data is often unstructured and does not convey meaningful information. Consequently, many data analysis methods have arisen. For instance, methods like deep learning can glean meaningful information from images by labelling them. In particular, the process of managing image data and labelling it is called image classification, and in some cases, it can surpass human-level accuracy. The computing system used to classify images is called a neural network, and the process through which the neural network learns to label the pictures is called image classification. Training a neural network is a process that generally requires a copious quantity of data to perform satisfactorily. Unfortunately, this requirement leads to training times for neural networks that are pretty long, and the availability of such an amount of data is often an issue. Transfer learning addresses some of these obstacles. Transfer learning is a technique within the machine learning field whose objective is improving the accuracy of a given task via pretraining while reducing the amount of data needed. This pretraining step precedes the actual task and entails training a neural network on a similar assignment on an unrelated domain. For example, in the context of image classification, the pretraining step could consist of training a model on a generic image dataset such as ImageNet, while the training proper is a task where the network must recognise several different races of cats and dogs. The neural network can perform better on the pets dataset because of the visual features learnt in the previous phase. However, the pretraining step still requires a dataset of a considerable dimension; therefore, the computational cost of creating a neural network remains a problem. This thesis introduces a novel way of optimising data efficiency in transfer learning, particularly the training time of image classification tasks. The proposed approach selects a subset of images from the source dataset via a semantic process, thus reducing the extent of data needed. For example, if the target dataset is dogs, the algorithm will only select those source dataset categories that are semantically close to canines. Our approach aids the neural network training process by reducing the number of photographs needed to create an image classifier. The main benefit is that it can assist machine learning researchers when developing new architectures. In addition, pretraining is often the most time and resource-consuming phase of training a machine learning model; therefore, researchers with time or energy constraints can benefit from this thesis's insights. The aforementioned semantic approach uses a lexical database. This database chooses the most relevant source dataset categories by determining the semantic distance between classes and selecting the target categories closest to the source classes.