Department
Computer Science
Major
Computer Science
Research Advisor
Tripathy, Ardhendu
Advisor's Department
Computer Science
Abstract
Active learning is a framework for training machine learning models where the goal is to reduce the number of labels used during training. Neural network models used for image classification require a large training dataset to achieve good accuracy. This project will use active learning to reduce the number of labels needed for training neural network models. We propose to use the Fisher information of the neural network parameters to actively select which images are labelled and included in the training data. A key challenge is the large number of parameters in commonly used neural network models, which significantly increases the cost of computing the Fisher information. To mitigate this challenge, we plan to identify a subset of the parameters that can be considered to be more relevant for the classification task. This project will obtain theoretical justification for our approach and implement it in the CIFAR1 O benchmark dataset.
Biography
Joshua Caruso is a Junior and National Merit Scholar working towards a F2024 graduation with a bachelor's in computer science and a minor in Mathematics. He has worked under Professor Tripathy as a research assistant for 2. 5 years and for the past year has helped design and implement strategies to increase labelling efficiency in active training of neural network models.
Research Category
Sciences
Presentation Type
Poster Presentation
Document Type
Poster
Award
Sciences Poster Session - Third Place
Location
Innovation Forum - 1st Floor Innovation Lab
Presentation Date
10 April 2024, 1:00 pm - 4:00 pm
A Fisher Information-based approach to improve labeling efficiency of neural network models in image classification
Innovation Forum - 1st Floor Innovation Lab
Active learning is a framework for training machine learning models where the goal is to reduce the number of labels used during training. Neural network models used for image classification require a large training dataset to achieve good accuracy. This project will use active learning to reduce the number of labels needed for training neural network models. We propose to use the Fisher information of the neural network parameters to actively select which images are labelled and included in the training data. A key challenge is the large number of parameters in commonly used neural network models, which significantly increases the cost of computing the Fisher information. To mitigate this challenge, we plan to identify a subset of the parameters that can be considered to be more relevant for the classification task. This project will obtain theoretical justification for our approach and implement it in the CIFAR1 O benchmark dataset.
