University of Twente Student Theses


Super Paretos: Bayesian Active Meta Learning for Spatial Transferability of Deep Learning Models

Sastry, Srikumar (2022) Super Paretos: Bayesian Active Meta Learning for Spatial Transferability of Deep Learning Models.

[img] PDF
Abstract:A common objective to address in the current realm of artificial intelligence (AI) is achieving more with less. This follows the notion of pareto which attributes 80\% of results to just 20\% of inputs. With advancing technology and research, it has become possible for AI to achieve super pareto, which further widens the ratio between necessary inputs and corresponding expected results. Keeping this in mind, we can draw direct correspondences between paretos and data efficiency. Data inadequacy has been a common problem in modern deep learning applications. It becomes a serious challenge especially in remote sensing applications where data collection and annotation is time consuming and expensive. A plethora of methods have been proposed in the past that fall into some theme of semi-supervised learning, active learning or transfer learning. These methods promise to solve the problem of data inadequacy and uphold the super pareto notion. However, a simple survey of these methods shall highlight their limited practical applicability on real life datasets. Such methods not only are computationally expensive but also fail to perform in varying realistic settings. In this study, we focus on one of the themes, namely active learning, commonly used to address the issue of efficient data assimilation. We take a holistic approach and describe its possible applications in imaging science. Most studies fail to uphold the chief principle that active learning was built on. We elaborate these limitations further in this study and propose novel active learning methods for image classification and segmentation. We not only compare our method with existing baselines but also present their results on real life remotely sensed datasets. One of the applications we focus on is crop mapping. Automatic crop mapping is important to ensure food security and efficient crop management. With a traditional deep learning approach, we are able to achieve an accuracy of 79.34\%. With active learning, we are able to achieve a similar result but with only 1.31\% of the data previously used. This translates to reducing about 75x efforts required for data collection and annotation. We see that this gap between necessary inputs and results is especially noticeable in practical datasets and therefore believe the notion of super pareto to become the new normal. Further in the study, we present the online learning setup for few shot learning which is extremely common in the real world. To this end, we propose an active meta learning method to understand the advantages of meta learning over offline/batch learning.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:54 computer science
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page