Parameter Reduction on Deep Learning Models
41Parameter Reduction (%)Real Block-Circ Trans
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Real Block-Circ TransStructured Representation=Real block-circ + DCT-DST, Model Component=Attention / FFN, Reduction Approach=Use sparse weight matrices instead of full ones.2026.05 | 41 | — | — | — | — | — | — | — | |
| Block-Circ AdaptersStructured Representation=Block-circulant adapters, Model Component=LLM fine-tuning, Reduction Approach=Compress only adapter modules.2026.05 | 14 | — | — | — | — | — | — | — | |
| CirCNNStructured Representation=Block-circulant, Model Component=CNN / FC, Reduction Approach=Combine all the blocks into a single vector and perform an FFT.2026.05 | — | 2 | — | — | — | — | — | — | |
| LDR TransformsStructured Representation=Low-displacement-rank, Model Component=Edge ML models, Reduction Approach=Weight sharing through fast transforms2026.05 | — | — | — | — | 3.5 | — | — | — |