Table 4 Detailed search cost and classification performance comparison for one-shot NAS methods on the Gastric Cancer datasets

From: Large language models driven neural architecture search for universal and lightweight disease diagnosis on histopathology slide images

Dataset: Gastric Cancer

 

ShuffleNet backbone

ViT backbone

Metric

Random search

Cream

Pathology-NAS

Random search

AutoFormer

Pathology-NAS

Iterations ↓

500

120

10

500

300

10

GPT-4 API Calls

0

0

10

0

0

10

FLOPs ↓

286.16M

430.04M

259.14M

4.25G

4.82G

4.25G

Prec@1 (%) ↑

62.47 ± 0.21

54.88 ± 0.28

63.15 ± 0.25***

40.62 ± 0.30

41.40 ± 0.12

43.04 ± 0.23***

Prec@5 (%)

98.61 ± 0.25

98.05 ± 0.31

98.99 ± 0.31**

93.57 ± 0.33

93.75 ± 0.25

94.28 ± 0.02***

API Cost ($)

0.00

0.00

0.15

0.00

0.00

0.16

Latency (hrs)

0.0000

0.0000

0.0011

0.0000

0.0000

0.0010

ST (GPU hrs) ↓

9.16

8.46

3.00

111.10

70.96

8.02

TT (GPU hrs) ↓

9.160

8.460

3.001

111.100

70.960

8.021

  1. Results are averaged over 5 independent runs. The table presents a comprehensive breakdown of search costs (Iterations, GPT-4 API Calls, Latency (hrs), API Cost ($), ST (GPU hrs), TT (GPU hrs)) alongside key performance metrics (FLOPs, Prec@1 (%), Prec@5 (%)). Optimal values for performance and lower values for costs are typically highlighted in bold where applicable. Statistical significance of Pathology-NAS Prec@1 (%) and Prec@5 (%) performance compared to Random Search (assessed by independent two-sample Welch’s t tests) is denoted by: **p < 0.01 (highly significant), ***p < 0.001 (very highly significant).
  2. ST Search Time, TT Total Time, TT = ST + Latency, Prec@1 Top-1 accuracy, Prec@5 Top-5 accuracy.
Back to article page