Image Classification on MNIST (Accuracy, Loss, and Resource Usage)
96.6AccuracyFull DNN
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Full DNN2026.07 | 96.6 | 0.1994 | 1,024 | 0 | 400,906 | 0.0004 | |
| Projected A, reachabilityTest=reachability2026.07 | 96.6 | 0.1996 | 998 | 2.54 | 657,418 | 0.0009 | |
| Projected B, observabilityTest=observability2026.07 | 96.6 | 0.1994 | 661 | 35.45 | 571,146 | 0.0009 | |
| Dynamic INT8 quantisation2026.07 | 96.5 | 0.2012 | 1,024 | 0 | 411,303 | 0.4466 | |
| Unstructured pruningConfig=50%2026.07 | 96.1 | 0.1805 | 1,024 | 0 | 200,970 | 0.0004 | |
| Projected C, balancedTest=balanced2026.07 | 96 | 0.2539 | 277 | 72.95 | 472,842 | 0.0009 | |
| Realised C-balanced2026.07 | 95.45 | 0.237 | 277 | 72.95 | 106,323 | 0.0004 | |
| Distilled realised C-balancedTraining=Distillation2026.07 | 94.7 | 0.2455 | 277 | 72.95 | 106,323 | 0.0004 | |
| Structured pruning + fine-tuning2026.07 | 91.75 | 0.3186 | 277 | 72.95 | 106,323 | 0.0004 | |
| Linear classifier2026.07 | 90.7 | 0.3305 | 0 | — | 7,850 | 0.0001 | |
| Low-rank SVDRank fraction=0.52026.07 | 85.5 | 0.5079 | 1,024 | 0 | 332,092 | 0.0007 | |
| Structured pruningTarget=C ranks2026.07 | 34.5 | 2.2544 | 277 | 72.95 | 106,323 | 0.0004 |