Research Topic
Sparse Neural Networks and Sparse Linear Algebra
Supervisor
- Professor Achim Streit (Karlsruhe Institute of Technology)
Thesis Abstract
Machine learning relies heavily on matrix/tensor operations. While dense representations dominate (e.g., transformers), sparse methods are growing (spiking/graph neural networks, pruned weights), offering regularization, better generalization, and efficiency gains. Existing Python libraries offer limited sparse support. Ginkgo, a C++ sparse linear algebra library, is being extended via pyGinkgo (Python bindings) with deep learning integration. Our work includes sparse convolution kernels and distributed SpMV/solvers in Ginkgo.
Publications
- Keshvi Tuteja et al. pyGinkgo: A Sparse Linear Algebra Operator Frame work for Python. Oct. 2025. doi: 10.48550/arXiv.2510.08230.
Conference Presentations
- ICPP 2025, HAICON 2025
Scientific Awards
- ISC WHPC Poster Award 2025