Empowering Limited Data Domains: Innovative Transfer Learning Strategies for Small Labeled Datasets
Navaneetha Krishnan S, Mohnish Karthikeyan B, Shriman K. Arun, Christopher Columbus C · 2024
Traditional Indian Medicine faces the risk of extinction due to insufficient transfer of knowledge. This research highlights the critical need for preservation through modern technology and techniques. It addresses the challenges of limited labeled data by employing pioneering transfer learning approaches for classifying Indian Medicinal Plants, specifically leaves. The study utilizes diverse trained models and a bespoke model tailored for this task, systematically investigating transfer learning dynamics to mitigate issues such as overfitting and underfitting in neural network training. Rooted in Artificial General Intelligence (AGI) concepts, the research demonstrates how transfer learning enhances neural network efficiency, especially when labeled data is scarce. Practical application is illustrated with a dataset of plant images, providing a real-world context for evaluating transfer learning strategies. By leveraging models trained with large data in relevant domains, the research facilitates effective feature extraction and generalization despite limited labeled data. A significant contribution of the research is the comparison of outcomes from various models with the proposed custom CNN model, HerboraNet, which achieves a remarkable accuracy of 95%. This research underscores the superiority of transfer learning over training models with limited input, particularly in data-constrained domains, highlighting its pivotal role in enhancing model efficiency and adaptability.