IEEE Computational Intelligence Magazine - May 2023 - 90

better. OnCIFAR-100, although the accuracy
ofEX-Net is not as high, the number
of parameters in EX-Net only
consumes 16.8% and 12.5% of the numbers
of parameters in DenseNet-B
(k ¼ 40) and ResNeXt-29 (8x64d),
respectively, demonstrating a significant
reduction. EX-Net uses only 1/8 of the
GPU resources consumed by ResNeXt29
(8x64d). In summary, EX-Net achieves
higher accuracy than the state-ofthe-art
manually-designed network architectures
and significantly outperforms all
competing approaches on CIFAR-10.
Additionally, EX-Net employs substantially
fewer parameters than these comparison
algorithms.
2) Comparison With Semiautomatic
NASMethods
Regarding the semiautomatic NAS algorithms,
EX-Net outperforms Hierarchical
Evolution, Block-QNN-S, and ENAS
(macro) in terms of the test accuracy and
number of parameters while significantly
reducing the search time cost
(16x4500x reductions). Compared to
NASNet-A,EX-Net shows slightly worse
test accuracy but requires fewer parameters
than NASNet-A. Moreover, EX-Net is
100,000x faster thanNASNet-A and consumes
only approximately 1/500 of the
GPU resources consumed by NASNet-A.
DARTS and ENAS (micro) achieve
slightly better accuracy on CIFAR-10
than EX-Net, while EX-Net has fewer
parameters. DARTS and ENAS (micro)
have larger deviations in test accuracy than
EX-Net. With the same GPU resource
consumption, EX-Net takes 75x and 25x
less search time than these two methods.
Compared to P-DARTS and PCDARTS,
EX-Net consumes fewer
parameters and GPU Days while showing
slightly reduced test accuracy. Moreover,
EX-Net can achieve more minor deviationsacrossdifferent
search rounds.SPDARTS
achieves slightly better accuracy
on CIFAR-10 than EX-Net but has
nearly 2x more parameters and 5x more
search time. In addition, although the
accuracy of EX-Net is less than that of
TE-NAS, the number ofEX-Net parameters
and the number of GPU Days
consumed by EX-Net are only half of
TABLE II Search time (minutes) on CIFAR-10 and CIFAR-100 by LoNAS.
NO. OF RUNS 1 2 3 4 5 6 7 8 9 10 AVERAGE VARIANCE
CIFAR-10
28.8 29.5 28.4 28.7 29.4 29.1 28.7 28.3 29.4 28.8 28.91
CIFAR-100
28.7 28.6 29.4 29.3 29.1 28.6 28.9 29.4 29.3 29.4 29.07
0.16
0.10
TABLE III The best architecture on CIFAR-10 by comprehensively considering the test
accuracy and the number of parameters of the network architectures.
TYPE
Conv Unit
Reg Unit1
BLOCK
-
Reg Block1
Reg Block2
Reg Block3
Reg Block4
Reg Unit2
Reg Block5
Reg Block6
Reg Block7
Reg Unit3
Reg Block8
Reg Block9
Reg Block10
Reg Block11
CONFIGURATION
input size=32*32, input channel=3, output channel=64
group=16, width=8, hasSENet=1, f=1
group=16, width=32, hasSENet=1, f=1
group=8, width=8, hasSENet=1, f=2
group=64, width=16, hasSENet=1, f=2
group=8, width=8, hasSENet=0, f=1
group=16, width=8, hasSENet=1, f=1
group=32, width=8, hasSENet=1, f=1
group=4, width=16, hasSENet=1, f=1
group=64, width=8, hasSENet=1, f=2
group=32, width=4, hasSENet=1, f=1
group=32, width=16, hasSENet=0, f=1
TABLE IV The best architecture on CIFAR-100 by comprehensively considering the test
accuracy and the number of parameters of the network architectures.
TYPE
Conv Unit
Reg Unit1
BLOCK
-
Reg Block1
Reg Block2
Reg Block3
Reg Unit2
Reg Block4
Reg Block5
Reg Block6
Reg Block7
Reg Unit3
Reg Block8
Reg Block9
Reg Block10
Reg Block11
Reg Unit4
Reg Block12
Reg Block13
Reg Block14
Reg Block15
those required by TE-NAS. Compared to
AS-NAS, EX-Net shows lower test
accuracy. However, EX-Net consumes
nearly 9x fewer parameters and requires
90 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2023
CONFIGURATION
input size=32*32, input channel=3, output channel=64
group=8, width=32, hasSENet=1, f=1
group=16, width=4, hasSENet=1, f=1
group=4, width=32, hasSENet=0, f=1
group=32, width=8, hasSENet=1, f=2
group=16, width=16, hasSENet=1, f=1
group=16, width=4, hasSENet=0, f=1
group=16, width=4, hasSENet=1, f=1
group=4, width=16, hasSENet=0, f=1
group=4, width=16, hasSENet=0, f=1
group=16, width=16, hasSENet=0, f=1
group=64, width=4, hasSENet=1, f=1
group=16, width=8, hasSENet=1, f=2
group=16, width=8, hasSENet=1, f=1
group=16, width=16, hasSENet=1, f=1
group=32, width=16, hasSENet=1, f=1
only 8x fewer computational resources
than AS-NAS and is thus competitive
algorithm. Therefore, compared to the
semiautomatic NAS algorithms, EX-Net
IEEE Computational Intelligence Magazine - May 2023

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2023
