Activity
Time:
Defense: Atharva Tendle
Date:
10:00 am
Zoom
“’The Revolution Will Not Be Supervised’: An Investigation of The Efficacy and Reasoning Process of Self-Supervised Representations”
Abstract: Learning visual representations without human supervision is a challenging problem. Recent advancements in the self-supervised learning (SSL) approach made it possible to learn effective representations from unlabeled data for solving computer vision tasks such as object recognition and detection. The performance of the Deep Learning image classification models trained using SSL representations is on par with the models based on representations learned via the state-of-the-art supervised learning (SL) techniques. However, no study has been done to determine the efficacy of SSL representations on various data domains as well as to understand the science of its efficacy. In this thesis, we conduct a multi-dimensional investigation on the SSL approach for performing classification. We identify the space of SSL’s excellence by investigating various SSL techniques on two types of target datasets: target dataset that is similar to the source dataset used to create the representations, and target dataset that is significantly different from the source dataset. For the latter type, we use the camera trap dataset that assembles various information on wildlife populations. In addition to this, we explain the effectiveness of SSL representations by two techniques: group symmetry-based analysis (e.g., invariance to various transformations) and feature map-based analysis. We design and conduct an extensive study for this investigation. The main contribution of this thesis is three-fold: (i) We achieve the new state-of-the-art benchmark on a standard and large camera trap dataset using the SSL-based approach. (ii) We analyze the effectiveness and efficiency of the main SSL techniques against the dominant SL technique for diverse domains. (iii) We provide a framework to study the group symmetries and interpretability of the SSL representations. Using this framework, we offer insights on how SSL models reason about the semantic identity of the target data used in a classification task.
Committee:
Dr. Mohammad Hasan, Advisor
Dr. Stephen Scott
Dr. Andrew Little
Join Zoom Meeting: https://unl.zoom.us/j/92015881814
Meeting ID: 920 1588 1814
Abstract: Learning visual representations without human supervision is a challenging problem. Recent advancements in the self-supervised learning (SSL) approach made it possible to learn effective representations from unlabeled data for solving computer vision tasks such as object recognition and detection. The performance of the Deep Learning image classification models trained using SSL representations is on par with the models based on representations learned via the state-of-the-art supervised learning (SL) techniques. However, no study has been done to determine the efficacy of SSL representations on various data domains as well as to understand the science of its efficacy. In this thesis, we conduct a multi-dimensional investigation on the SSL approach for performing classification. We identify the space of SSL’s excellence by investigating various SSL techniques on two types of target datasets: target dataset that is similar to the source dataset used to create the representations, and target dataset that is significantly different from the source dataset. For the latter type, we use the camera trap dataset that assembles various information on wildlife populations. In addition to this, we explain the effectiveness of SSL representations by two techniques: group symmetry-based analysis (e.g., invariance to various transformations) and feature map-based analysis. We design and conduct an extensive study for this investigation. The main contribution of this thesis is three-fold: (i) We achieve the new state-of-the-art benchmark on a standard and large camera trap dataset using the SSL-based approach. (ii) We analyze the effectiveness and efficiency of the main SSL techniques against the dominant SL technique for diverse domains. (iii) We provide a framework to study the group symmetries and interpretability of the SSL representations. Using this framework, we offer insights on how SSL models reason about the semantic identity of the target data used in a classification task.
Committee:
Dr. Mohammad Hasan, Advisor
Dr. Stephen Scott
Dr. Andrew Little
Join Zoom Meeting: https://unl.zoom.us/j/92015881814
Meeting ID: 920 1588 1814