Autotuned Classification Based on Knowledge Transferred from Self supervised Models

Abstract

This thesis focuses on improving the efficacy of deep self-supervised CNN models newlinefor image classification over the target datasets. With the recent advancements of deep newlinelearning-based methods in supervised image classification, the requirement of a huge newlineamount of labeled training data is inevitable to avoid overfitting problems. Preparing newlinesuch a huge amount of labeled data takes a lot of human effort and time. In this newlinescenario, self-supervised models are becoming popular because of their ability to learn newlineeven from unlabeled datasets. However, the efficient transfer of knowledge learned by newlineself-supervised models into a target task is an unsolved problem. newlineFirst, this thesis proposes a method for the efficient transfer of knowledge learned by newlinea self-supervised model, into a target task. The hyperparameters such as the number of newlinelayers, number of units in each layer, learning rate, and dropout are automatically tuned in newlinethe Fully Connected (FC) layers, using a Bayesian optimization technique called the Treestructured newlineParzen Estimator Approach (TPE) algorithm. To evaluate the performance of newlinethe proposed method, state-of-the-art self-supervised models such as SimClr and SWAV newlineare used to extract the learned features. newlineSecond, we extended the concept of automatically tuning (autotuning) the newlinehyperparameters (as proposed in the first problem), to apply on CNN layers. This work newlineuses a pre-trained self-supervised model to transfer knowledge for image classification. newlineWe propose an efficient Bayesian optimization-based method for autotuning the newlinehyperparameters of the self-supervised model during the knowledge transfer. The newlineproposed autotuned image classifier consists of a few CNN layers (the number of layers newlineis learned) followed by an FC layer. Finally, we use a softmax layer to obtain the newlineprobability scores of classes for the images. newlineThird, we further focus on parameter overhead and GPU usage for hours. This newlinework proposes a method to address the two significant challenges for an image newlineclassification task: the labeled datasets