Optimizing CNN memory efficiency: a continual learning solution for plant stress classification

Abstract

In nature, plants encounter a wide variety of stress-inducing elements that can create unfavorable conditions or impede their metabolic processes and hinder growth. Both biotic, living organisms, and abiotic, non-living elements, can contribute to this stress in plants. Bangladesh, being an agricultural nation, has a lot of difficulties in detecting symptoms of these stresses. Our study focuses on recognizing and classifying stress in four popular plants that are consumed both domestically and internationally, including eggplant, bitter gourd, snake gourd and ash gourd. In recent times, Convolutional Neural Network (CNN) is widely and successfully used to classify health conditions from leaf images. But the CNN models in use today are vulnerable to a phenomenon known as catastrophic forgetting. Because of this, currently, there is no CNN model that can perform well in several tasks at once, without forgetting the previous tasks. This thesis is to empower a CNN model with Continual Learning (CL), that can learn new tasks continuously without forgetting its prior knowledge. In this study, we evaluate and demonstrate the issue of forgetting prior knowledge, that can be eradicated by using Continual Learning (CL) approaches like Elastic Weight Consolidation (EWC) and Learning without Forgetting (LwF), while being able to perform multi-task and achieving significantly better performance than conventional approach. To simulate a scenario similar to continuous data flow, the dataset is divided into four different tasks. Where without the Continual Learning (CL) methods the CNN model EfficientNet-B0 suffers from catastrophic forgetting achieving accuracy of 42%, 55%, 67%, and 97% for task-1, task-2, task-3, and task-4 respectively after training on task-4. After applying Continual Learning (CL) a drastic improvement of 7% to 18% in accuracy was achieved throughout different tasks. Elastic Weight Consolidation (EWC) achieves 60%, 65%, 66%, and 98% for task-1, task-2, task-3, and task-4 respectively after training on task-4 . Learning without Forgetting (LwF) achieves 60%, 61%, 55%, and 75% for task-1, task-2, task-3, and task-4 respectively after training on task-4. As Continual Learning (CL) approach preserves the knowledge of previous tasks while only training on the current task, this method can significantly reduce the training time and memory issue of CNN models while training on multiple tasks.