Analisis Komparatif Kinerja dan Stabilitas Optimizer Adam, AdamW, dan Lion pada Convolutional Neural Network ConvNeXt-Tiny Menggunakan Klasifikasi Gambar CIFAR-100

Purba, Taty Eirene Vyalatama (2026) Analisis Komparatif Kinerja dan Stabilitas Optimizer Adam, AdamW, dan Lion pada Convolutional Neural Network ConvNeXt-Tiny Menggunakan Klasifikasi Gambar CIFAR-100. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221065-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Penelitian ini membandingkan performa klasifikasi, kecepatan konvergensi, stabilitas numerik, waktu komputasi, dan penggunaan memori GPU dari optimizer Adam, AdamW, dan Lion pada arsitektur ConvNeXt-Tiny untuk klasifikasi citra CIFAR-100. Model dilatih dari awal melalui dua skenario eksperimen. Pada skenario pertama, setiap optimizer diuji menggunakan tiga random seed selama maksimum 100 epoch. Pada skenario kedua, stabilitas optimizer dianalisis menggunakan variasi learning rate dari 10⁻⁵ hingga 10⁻² selama 20 epoch. Hasil menunjukkan bahwa AdamW memberikan performa klasifikasi terbaik dengan rata-rata Top-1 Accuracy 54,05%, Top-5 Accuracy 78,42%, dan validation loss 3,01. Lion berada pada posisi kedua dengan Top-1 Accuracy 50,35%, sedangkan Adam menghasilkan Top-1 Accuracy terendah sebesar 26,15%. AdamW juga mencapai 95% dari akurasi validasi terbaik paling cepat, dengan median 57 epoch, diikuti Adam 59 epoch dan Lion 61 epoch. Namun, Lion menunjukkan konsistensi konvergensi terbaik dengan standar deviasi 1,00. Pada pengujian stabilitas numerik, Lion mencapai akurasi 43,8% pada learning rate 10⁻³, tetapi menurun tajam menjadi 19,4% pada 10⁻², yang menunjukkan sensitivitas lebih tinggi terhadap perubahan learning rate. AdamW mencatat penggunaan memori GPU terendah sebesar 667 MB, sedangkan Adam memiliki waktu komputasi tercepat sebesar 52,92 detik per epoch. Uji Kruskal-Wallis menghasilkan H=7,2 dan p=0,0273, sehingga terdapat perbedaan keseluruhan yang signifikan pada Top-1 Accuracy. Secara keseluruhan, AdamW memberikan keseimbangan terbaik antara akurasi, kecepatan konvergensi, stabilitas, dan efisiensi komputasi pada konfigurasi penelitian ini.
===========================================================================================================================
This study compares the classification performance, convergence speed, numerical stability, computation time, and GPU memory usage of Adam, AdamW, and Lion on ConvNeXt-Tiny for CIFAR-100 image classification. The model was trained from scratch under two experimental scenarios. In the first scenario, each optimizer was evaluated across three random seeds for up to 100 epochs. In the second scenario, optimizer stability was examined using learning rates from 10⁻⁵ to 10⁻² over 20 epochs. AdamW achieved the best overall classification performance, with a mean Top-1 accuracy of 54.05%, Top-5 accuracy of 78.42%, and validation loss of 3.01. Lion ranked second with a Top-1 accuracy of 50.35%, while Adam produced the lowest Top-1 accuracy of 26.15%. AdamW also reached 95% of its best validation accuracy fastest, with a median of 57 epochs, followed by Adam at 59 epochs and Lion at 61 epochs. However, Lion showed the most consistent convergence across seeds, with a standard deviation of 1.00. In numerical stability testing, Lion achieved 43.8% accuracy at a learning rate of 10⁻³ but declined sharply to 19.4% at 10⁻², indicating greater sensitivity to learning-rate changes. AdamW recorded the lowest GPU memory usage at 667 MB, whereas Adam had the shortest computation time at 52.92 seconds per epoch. The Kruskal-Wallis test yielded H=7.2 and p=0.0273, indicating a significant overall difference in Top-1 accuracy. Overall, AdamW provided the best balance of accuracy, convergence speed, stability, and computational efficiency for the evaluated ConvNeXt-Tiny and CIFAR-100 configuration, although these findings remain specific to the experimental setup used.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Adam, AdamW, ConvNeXt-Tiny, Green AI, Lion, optimizer, CIFAR-100, Adam, AdamW, ConvNeXt-Tiny, Green AI, Lion, optimasi, CIFAR-100
Subjects:	Q Science > QA Mathematics > QA9.58 Algorithms
Divisions:	Faculty of Mathematics and Science > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Taty Eirene Vyalatama. P
Date Deposited:	03 Aug 2026 03:20
Last Modified:	03 Aug 2026 03:20
URI:	http://repository.its.ac.id/id/eprint/141589

Actions (login required)

View Item