Learning with Limited Datasets: From Deep-Learning to Traditional Machine-Learning

Learning with datasets containing a limited number of exemplars is a contentious research area. Researchers have used deep learning (DL) models with a large number of trainable parameters for such limited datasets leading to problems such as overfitting, over-parameterization, lack of generalization, and the need for large computational resources. Judicious use of appropriate learning methodologies may be in order when the dataset for training is limited. Non-DL methodologies or traditional machine learning methodologies with appropriate pre-processing and feature extraction techniques may perform at par or better than DL techniques for applications that have a limited dataset. This dissertation aims to establish this proposition for two healthcare applications namely, radar-based monitoring of human activities (and fall event detection) and thermography-based breast abnormality detection by developing computationally inexpensive novel supervised and unsupervised non-DL learning methodologies for binary and multi-class classification problems that outperform the current state-of-the-art techniques of the respective fields in those healthcare applications. The developed learning methodologies use traditional machine learning classifiers along with interpretable hand-crafted features such as histograms of oriented gradient (HOG), statistical features, and textural features. These novel learning methodologies use ensemble learning approaches such as early fusion, intermediate fusion, decision fusion, or training error correction. Novel contrast enhancement and novel gradient enhancement methodologies using binary encodings such as census transform and local binary patterns are also proposed to improve classification performance. For both applications, learning in the compressed domain using deterministic compressive sensing is introduced to reduce the number of trainable parameters of the developed novel supervised non-DL methodologies. The novel supervised learning methodologies using hand-crafted features achieved an average accuracy of 98% and 96% for fall event detection and breast abnormality detection, respectively. The novel unsupervised learning methodologies using hand-crafted features achieved an average error rate of 1.1% for fall event detection and an average F1-score of 85% for breast abnormality detection. At 0.875 compression ratio, the novel supervised learning methodologies in the compressed domain achieved an average accuracy of 97% and 87% for fall event detection and breast abnormality detection, respectively.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.