인문학
사회과학
자연과학
공학
의약학
농수해양학
예술체육학
복합학
DBpia AI
AI 에이전트
AI 아이디어
AI 리더
AI 뷰어
Citeasy
크롬 서지관리 다운로드
지원사업
학술연구/단체지원/교육 등 연구자 활동을 지속하도록 DBpia가 지원하고 있어요.
커뮤니티
연구자들이 자신의 연구와 전문성을 널리 알리고, 새로운 협력의 기회를 만들 수 있는 네트워킹 공간이에요.
논문 기본 정보
- 저자정보
피인용 0
초록·키워드
With the recent significant improvement in the performance of Transformer-based language models, the importance of explainability regarding the internal processes that generate prediction results has increased. However, relying solely on observational metrics such as attention weights to determine whether specific components contribute to actual predictions has limitations. Accordingly, this study constructed a small-scale Decoder-only Transformer and analyzed layer-by-layer internal representations, attention, ablation, and gradients. The model was trained using data on capitals, currencies, and continental relationships, and the reproducibility of the results was verified through repeated experiments with five different seeds. The experimental results showed that while the internal representation of the ground truth token became increasingly clear as the layers deepened, the most significant performance degradation was consistently observed when the initial layer was removed during layer ablation. Furthermore, although the number of the most important attention head varied depending on the seed, it appeared in Layer 1 in all cases. This demonstrates that the explicitness of information observed in internal representations does not necessarily correlate with the functional importance of actual predictions. This study proposes a method to analyze the internal prediction mechanism of Transformers by combining attention visualization with ablation-based interventions.
본문·목차
인공지능 문자 인식 모델을 통해 추출된 텍스트로, 일부 오타나 오류가 포함될 수 있으나 지속적으로 개선 중입니다.
오류를 발견하셨다면 해당 부분을 드래그한 후 ' 를 통해 신고해주세요.
오류를 발견하셨다면 해당 부분을 드래그한 후 ' 를 통해 신고해주세요.
최근 본 자료 전체보기
