Spatial Sparse PCA Using Autoencoders with Convolutions
Testing whether convolution can preserve local spatial structure in sparse learned representations.
Research
Thesis research
Explanation Analysis Using Rule Extraction
Generating faithful, human-readable rules around black-box model predictions.
Research threads
01
Structured representation learning
How can inductive biases such as convolution help sparse latent representations preserve local spatial relationships? My statistics thesis explored this question through simulation and climate data.
02
Trustworthy & explainable AI
How can a black-box prediction be translated into a compact set of rules without losing fidelity? At Ericsson R&D, I developed and evaluated a post-hoc, model-agnostic rule extraction pipeline.
03
Self-supervised & predictive learning
I am interested in how models can discover correspondence and dynamics from unlabeled video, then use learned representations to anticipate future states rather than only reconstruct observations.