Overview
Lightweight Deep Learning with MST explored model compression through graph algorithms. The project represented model-weight relationships as a graph and used minimum-spanning-tree algorithms to reduce parameters while preserving useful structure.
Approach
- Created correlations between model weights and graph edges.
- Generated minimum spanning trees with Kruskal, Prim, and Floyd-style graph-processing experiments.
- Compared compression behavior against model-performance retention.
- Implemented and evaluated the experiments with PyTorch.
Result
The experiment verified that the model could be compressed by about 50% while preserving roughly 70% of the original performance in the tested setting.
Takeaway
The project connected classical graph algorithms with neural-network compression, giving a practical view of how algorithmic structure can be used to reason about model size and inference cost.
Materials
This visual summarizes the graph-based model-compression idea: representing weight relationships as a graph and preserving a compact structure through minimum spanning tree selection.
