An active learning Approach based on learning models' parameters exploitation

Abstract

Arti cial Intelligence (AI) techniques and in particular Machine and Deep Learning (ML and DL), have been widely adopted to enhance various aspects of human life. ML algorithms can be categorized into four main types: Supervised Learning, Unsupervised Learning, Semi-supervised Learning, and Reinforcement Learning. A signi cant challenge in these techniques is the requirement for su cient labeled data for training. Active Learning (AL) is a machine learning framework that addresses this issue by selecting instances to be labeled in a smart way to optimize model training, i.e., AL reduces labeling time and leads to better-performing models by dynamically selecting the most representative samples to be labeled during the training phase. AL was proven to be e ective in di erent scenarios and its choice of querying a label depends on the cost and gain of obtaining the information. In this thesis, are presented two novel approaches for active learning in meta-learning models. The proposed methods, called LAL-IGradV and LAL-IGradV-VAE, select instances to be labeled using an estimate of their impact on the current classi er. This is achieved by evaluating the importance of previously labeled instances in training the classi cation model and training another model that estimates the importance of unlabeled instances. The approaches can be instantiated with any classi er that is trainable through gradient descent optimization, and in this study, is provided a formulation using a deep neural network. These approaches have not been thoroughly investigated in previous learning-to-active-learn methods and experimental results demonstrate its promising performance in scenarios where there are only a limited number of initially labeled instances. 2