My work focuses on tackling (or at least trying to 😅) some challenges related to machine learning and data analysis. I specialize in designing frameworks to handle messy, real-world data, particularly in environments filled with noise, variability, and non-stationarity. Lately, I've been applying these ideas to practical fields like precision agriculture. Here, you’ll find my research projects, experiments, and tools that aim to push the boundaries of conventional machine learning methods.
Supervised learning traditionally relies on expert labeling, but when dealing with large datasets, this can be both costly and time-consuming. I work on alternative methods to handle multiple annotators, with varying expertise levels, through Generalized Cross-Entropy-based Chained Deep Learning (GCECDL).
- Key Contributions:
- Developed a framework that adapts to each annotator's non-stationary patterns and preserves inter-dependencies between them.
- Combines deep learning with a noise-robust loss function to tackle unreliable labelers.
- Achieves robust classification performance, outperforming state-of-the-art methods.
Agricultural management and crop optimization require the analysis of complex datasets, often characterized by high variability and non-linear relationships. My work in Local Biplot provides insights into such data, focusing on crop water status prediction.
- Key Contributions:
- Introduced Local Biplot, a method combining UMAP with affine transformations and SVD decomposition to codify non-stationary data patterns.
- Applied the method to remote sensing data for precision agriculture, improving feature relevance and prediction accuracy.
- Enhanced interpretability and clustering performance for both synthetic and real-world datasets.
