Paper Accepted at the 28th European Conference on Artificial Intelligence (ECAI 2025)

Lee Tae-han, combined master's–doctoral student (supervisor: Professor Lee Hyuk-jun)

The paper 'Token Pruning in Audio Transformers: Optimizing Performance and Decoding Patch Importance', written by Lee Tae-han (combined master's–doctoral, first author) of the Intelligent Computer Architecture and Embedded Computing Laboratory (ECL) and Professor Lee Hyuk-jun (corresponding author), has been accepted for oral presentation and publication at the European Conference on Artificial Intelligence (ECAI 2025).

▶ The structure of a transformer block with the token pruning module applied

The research applies token pruning to transformer-based audio classification models, evaluating the importance of each token on the basis of attention scores, and reduces computation (MAC) at inference by 30–40% while keeping accuracy loss below about 1%. Quantitative analysis also verified that although the attention scores of tokens correlate quantitatively with the energy of the corresponding mel-spectrogram patch, tokens generated in low-energy regions still receive considerable attention scores compared with pruned tokens, and that their presence plays an essential role in audio classification.

The 28th European Conference on Artificial Intelligence (ECAI-2025) will be held in Bologna, Italy (25–30 October), and is listed among the outstanding international conferences in computer science under the BK21 Plus programme (BK IF=1).

References:

- Conference website: https://ecai2025.org/

- Paper: https://arxiv.org/abs/2504.01690