Metode Hybrid Consolidation-Replay Pada Task Incremental Learning Dalam Studi Kasus Klasifikasi Produk Fashion

Alamsyah, Ryan Badai (2026) Metode Hybrid Consolidation-Replay Pada Task Incremental Learning Dalam Studi Kasus Klasifikasi Produk Fashion. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241013-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Jaringan saraf dalam (Deep Neural Networks) modern unggul dalam klasifikasi citra statis, namun gagal beradaptasi di dunia nyata yang dinamis seperti industri fashion, di mana tren dan kategori produk baru muncul secara konstan. Ketika model yang ada dilatih pada data baru, model tersebut mengalami catastrophic forgetting, yaitu kegagalan fundamental di mana pengetahuan lama secara drastis terlupakan saat mempelajari informasi baru. Penelitian ini berfokus untuk mengatasi masalah tersebut dalam skenario Task Incremental Learning (TIL), di mana model dituntut untuk mempelajari tugas-tugas baru secara sekuensial tanpa melupakan tugas-tugas sebelumnya. Untuk mengatasi permasalahan tersebut, penelitian ini mengusulkan metode bernama Hybrid Consolidation-Replay (HCR), yang mensinergikan antara Elastic Weight Consolidation (EWC) dan Experience Replay (ER). EWC berfungsi sebagai metode regularisasi yang secara selektif melindungi parameter model yang krusial untuk tugas-tugas lama, sementara ER menyimpan sebagian kecil sampel data lama ke dalam buffer memori dan melatihnya kembali bersama data baru. Metode HCR diimplementasikan pada arsitektur backbone Inception-V3.
Efektivitas HCR dievaluasi pada dua dataset, yaitu Fashion-MNIST sebagai baseline akademis dan dataset in-the-wild dari platform e-commerce Carousell, masing-masing dengan skenario dua tugas dan empat tugas, serta dibandingkan dengan Sequential Fine-Tuning, Joint Training, EWC, dan ER. Hasil eksperimen menunjukkan bahwa HCR secara konsisten mencapai Average Accuracy (AA) tertinggi di antara seluruh metode continual learning pada keempat skenario pengujian (AA: 0.9934 pada empat tugas, 0.8533 pada dua tugas Carousell, dan 0.9540 pada dua tugas Fashion-MNIST), sekaligus mencatatkan Backward Transfer (BWT) terkecil yang mendekati nol dengan reduksi forgetting hingga 97.7% dibandingkan Sequential Fine-Tuning. HCR juga terbukti lebih efisien secara komputasi dibandingkan Joint Training maupun ER tunggal. Secara keseluruhan, HCR berhasil membuktikan bahwa sinergi EWC dan ER mampu mengungguli kedua metode tunggal sekaligus mendekati batas atas performa ideal dengan efisiensi komputasi yang jauh lebih baik
==================================================================================================================================
Modern deep neural networks excel at classifying static images, but fail to adapt to dynamic real-world environments such as the fashion industry, where new trends and product categories constantly emerge. When existing models are trained on new data, they experience catastrophic forgetting, a fundamental failure in which previously acquired knowledge is drastically forgotten while learning new information. This research focuses on addressing this issue in a Task Incremental Learning (TIL) scenario, where the model is required to learn new tasks sequentially without forgetting previous tasks.
To address this problem, this research proposes a method called Hybrid Consolidation-Replay (HCR), which combines Elastic Weight Consolidation (EWC) and Experience Replay (ER). EWC serves as a regularization method that selectively protects model parameters crucial for older tasks, while ER stores a small portion of old data samples in a memory buffer and retrains them alongside new data. The HCR method is implemented on the Inception-V3 backbone architecture.
The effectiveness of HCR was evaluated on two datasets: Fashion-MNIST as an academic baseline and an in-the-wild dataset from the Carousell e-commerce platform, each with two-task and four-task scenarios, and compared against Sequential Fine-Tuning, Joint Training, EWC, and ER. Experimental results show that HCR consistently achieves the highest Average Accuracy (AA) among all continual learning methods across all four test scenarios (AA: 0.9934 for the four-task scenario, 0.8533 for the two-task Carousell scenario, and 0.9540 for the two-task Fashion-MNIST scenario), while also recording the smallest Backward Transfer (BWT), nearing zero, with a reduction in forgetting of up to 97.7% compared to Sequential Fine-Tuning. HCR also proved to be more computationally efficient than both Joint Training and standalone ER. Overall, HCR successfully demonstrated that the synergy between EWC and ER can outperform both standalone methods while approaching the upper bound of ideal performance with significantly better computational efficiency.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Catastrophic Forgetting, Elastic Weight Consolidation, Experience Replay, Hybrid Consolidation-Replay, Task Incremental Learning, Catastrophic Forgetting, Elastic Weight Consolidation, Experience Replay, Hybrid Consolidation-Replay, Task Incremental Learning
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Ryan Badai Alamsyah
Date Deposited:	28 Jul 2026 08:04
Last Modified:	28 Jul 2026 08:04
URI:	http://repository.its.ac.id/id/eprint/138805

Actions (login required)

View Item