Skip to main content
Ko Lab

Research

Stephany works on making deep learning models small and fast enough to run where the data is produced, from a drone over a canola field to a transformer on an edge device.

Current project

Grassland Genomics for Greenhouse Gas Mitigation (GG4GHG)

Stephany works on Grassland Genomics for Greenhouse Gas Mitigation, a Genome Canada project estimating how much carbon is stored in soil across Canadian grasslands.

Canadian grasslands cover more than 16 million hectares and play an important role in carbon storage, biodiversity and greenhouse gas mitigation, but grassland loss and management practices affect how much carbon they hold. Soil organic carbon is a key indicator of soil health, showing how effectively grasslands retain carbon while supporting fertility, productivity and resilience.

  • Combines microbial sequencing and environmental data with machine learning and feature selection to predict soil organic carbon
  • Identifies the main biological and environmental drivers of carbon storage
  • Explores graph-based models to represent interactions among microbes, plants, soil and climate
Mean test R-squared against the number of top-N selected 16S features: XGBoost plateaus near 0.52 and random forest near 0.48, both rising steeply up to about 50 features.

Publications

Optimized deep learning for canola yield prediction on edge devices

S. Valarezo-Plaza, J. Torres-Tello, K. D. Singh, S. J. Shirtliffe, S. Deivalakshmi, et al. IEEE Transactions on AgriFood Electronics, 2(2), 436-444, 2024.

Workflow from drone data acquisition through a deep learning model to deployment on an edge device.
  • Predicts canola yield from drone-captured hyperspectral imagery across 150 bands and eight dates
  • Feature selection, pruning and quantization shrink the model until it runs on an edge device in the field

Optimized transformer models: pruning and quantization for the edge

M. H. Haider, S. Valarezo-Plaza, S. Muhsin, H. Zhang, S.-B. Ko. IEEE International Symposium on Circuits and Systems (ISCAS), 2024.

  • Tests whether pruning and quantization designed for CNNs carry over to transformers, whose computation patterns differ
  • Large compression gains while transformer accuracy holds

FFT-based deep learning for combustion instability prediction

S. Valarezo-Plaza, A. Erazo, M. H. Haider, J. Bae, P. Canteenwalla, S. Yun, et al. International Journal of Hydrogen Energy, 2026.

  • Works in the frequency domain to predict combustion instability efficiently
Diagram of a BiLSTM network with sequence input, fully connected layer and softmax output.