A Comparative Analysis of HPS and MMoE Models for Multi-Task Learning: Empirical Evaluation of Office-31 and NYU-V2 Datasets

Menlolo, Ariel Pratama (2026) A Comparative Analysis of HPS and MMoE Models for Multi-Task Learning: Empirical Evaluation of Office-31 and NYU-V2 Datasets. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025211194-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Keseimbangan antara representasi bersama (shared representations) dengan kebutuhan unik dari setiap tugas individu tetap menjadi tantangan utama dalam Multi-Task Learning (MTL). Hambatan signifikan yang sering muncul adalah negative transfer, di mana backbone bersama gagal menangkap granulasi fitur yang optimal untuk tugas-tugas yang saling bertentangan. Penelitian ini menyajikan analisis komparatif antara arsitektur Hard Parameter Sharing (HPS) dan Multi-Gate Mixture-of-Experts (MMoE), dengan menggunakan Single- Task Learning (STL) sebagai dasar pembanding (baseline). Penelitian ini mengevaluasi model- model tersebut pada dua tolok ukur (benchmark) yang berbeda: dataset Office-31 untuk adaptasi domain dan dataset NYU-V2 untuk tugas-tugas RGB-D dalam ruangan, termasuk segmentasi semantik dan estimasi kedalaman. Metodologi yang digunakan melibatkan penilaian kemampuan mekanisme gating dinamis pada MMoE dalam mengalokasikan expert bersama secara efektif dibandingkan dengan struktur kaku pada HPS. Hasil empiris menunjukkan bahwa meskipun STL tetap unggul untuk dataset kecil dengan korelasi tinggi seperti Office-31, MMoE mencapai performa keseluruhan terbaik pada dataset NYU-V2 yang lebih kompleks dengan peningkatan metrik relatif (%m) sebesar 1,094 terhadap baseline. Secara khusus, MMoE unggul dalam tugas berbasis geometri, dengan mengurangi jarak sudut rata-rata (mean angle distance) pada prediksi surface normal menjadi 23,02° dibandingkan 25,58° pada STL. Evaluasi ini mengonfirmasi bahwa MMoE berhasil memitigasi interferensi tugas dan sinyal spesifik tugas dengan menavigasi keseimbangan yang rumit antara pembelajaran bersama dan spesialisasi. Penelitian ini menyoroti pentingnya arsitektur modular dalam sistem visi komputer modern yang menangani data multi-modal yang kompleks
====================================================================================================================
Balancing shared representations with the unique requirements of individual tasks remains a primary challenge in Multi-Task Learning (MTL). A significant obstacle is negative transfer, where a shared backbone fails to capture the optimal feature granularity for conflicting tasks. This study presents a comparative analysis between Hard Parameter Sharing (HPS) and Multi- Gate Mixture-of-Experts (MMoE) architectures, using Single-Task Learning (STL) as a baseline. The research evaluates these models on two distinct benchmarks: the Office-31 dataset for domain adaptation and the NYU-V2 dataset for indoor RGB-D tasks, including semantic segmentation and depth estimation. The methodology involves assessing the capability of MMoE’s dynamic gating mechanism to allocate shared experts effectively compared to the rigid structure of HPS. The empirical results demonstrate that while STL remains superior for smaller, highly correlated datasets like Office-31, MMoE achieves the best overall performance on the more complex NYU-V2 dataset with a relative improvement (%m) of 1.094 over the baseline. Notably, MMoE excelled in geometry-driven tasks, reducing the mean angle distance in surface normal prediction to 23.02° compared to 25.58° in STL. The evaluation confirms that MMoE successfully mitigates task interference and task-specific signals by navigating the delicate balance between shared learning and specialisation. This research highlights the importance of modular architectures in modern computer vision systems dealing with complex, multi-modal data.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Multi-Task Learning, MMoE, Hard Parameter Sharing, Negative Transfer, Office-31, NYU-V2.
Subjects:	T Technology > T Technology (General) > T11 Technical writing. Scientific Writing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Ariel Pratama Menlolo
Date Deposited:	06 Feb 2026 02:42
Last Modified:	06 Feb 2026 02:42
URI:	http://repository.its.ac.id/id/eprint/132212

Actions (login required)

View Item