Title

Muscle-Guided Latent Representation Distillation for Robust Spinal-Based Motor Intention Decoding

Alternative Title

근육 유도 잠재 표현 증류를 통한 견고한 척추 기반 운동 의도 디코딩

DGIST Authors

Chanyu Moon ; Ji-Woong Choi

Advisor

최지웅

Issued Date

2026

Awarded Date

2026-08-01

Type

Thesis

Description

Brain-Machine Interface （BMI）, Electrospinogram （ESG）, Cross-Modal Learning, Contrastive Learning, Motor Intention Decoding.

Table Of Contents

1. Introduction 1
2. Background and Related Works 4
2.1 Neural Basis of Motor Control 4
2.2 Biosignal Modalities for Motor Intention Decoding 5
2.2.1 Peripheral Signal （EMG） 5
2.2.2 Cortical Signals （EEG and ECoG） 6
2.2.3 Spinal Signal （ESG） 7
2.3 Challenges in ESG-based Decoding 7
3. Methods and Materials 10
3.1 Problem Definition 10
3.2 Framework Overview 11
3.2.1 Encoder 12
3.2.2 Decoder 12
3.3 Cross-Modal Supervision Methods 12
3.3.1 Latent-level Methods 14
3.3.2 Output-level Methods 17
3.4 Single-Modality Methods 18
3.5 Dataset 19
3.6 Experimental Setup 20
3.7 Evaluation Metrics 20
3.7.1 Mean Squared Error （MSE） 20
3.7.2 Mean Absolute Error （MAE） 21
3.7.3 R2 （Coefficient of Determination） 21
3.7.4 Spearman Correlation 21
3.8 Implementation Details 22
4. Results 23
4.1 Single Modality Baseline 23
4.2 Comparison of Cross-Modal Supervision Methods 24
5. Conclusion 29
5.1 Summary 29
5.2 Discussion 29
5.2.1 Effectiveness of Contrastive Alignment 29
5.2.2 Limitations of Direct Latent Matching 30
5.2.3 Limitations of Output-Level Supervision 31
5.2.4 Remaining Performance Gap with EMG 32
5.2.5 Limitations and Future Work 32

URI

https://scholar.dgist.ac.kr/handle/20.500.11750/60828
http://dgist.dcollection.net/common/orgView/200001006867

DOI

10.22677/THESIS.200001006867

Degree

Master

Department

Artificial Intelligence Major

Publisher

DGIST

Show Full Item Record