APRIL
A replay-free domain incremental learning framework for respiratory sound classification, built on a frozen audio backbone and sequentially isolated low-rank experts, with a multi-level prototype router that adaptively selects experts.
Timeline (last 90 days)
APRIL is an exemplar-free domain-incremental learning framework based on a shared frozen BEATs backbone that sequentially inherits and isolates low-rank experts, so that adaptation to a new domain does not alter the old models.
The core of APRIL is a multi-level prototype router: it first retrieves the top two candidate experts, then fuses the domain-routing posterior with classification confidence, and dynamically weights the selection of the final model according to matching ambiguity.
APRIL freezes the BEATs backbone as a shared feature extractor, avoiding representation drift caused by fine-tuning and greatly reducing trainable parameters.
Across evaluations with five random seeds, APRIL led the compared continual learning methods on final-session average scores.
APRIL's final-session average scores on the ICBHI dataset were 64.4% for forward and 65.4% for reverse.
APRIL's final-session average scores on the SPRSound dataset were 83.1% for forward and 82.2% for reverse.
Under all tested protocols, APRIL's average forgetting rate was below 1.0%.
APRIL's replay-free nature aligns with the privacy requirements of medical scenarios, providing a feasible path for deploying continuous breath sound classification in real medical environments.
All Facts (8)
APRIL is an exemplar-free domain-incremental learning framework based on a shared frozen BEATs backbone that sequentially inherits and isolates low-rank experts, so that adaptation to a new domain does not alter the old models.
50%UnverifiedThe core of APRIL is a multi-level prototype router: it first retrieves the top two candidate experts, then fuses the domain-routing posterior with classification confidence, and dynamically weights the selection of the final model according to matching ambiguity.
50%UnverifiedAPRIL freezes the BEATs backbone as a shared feature extractor, avoiding representation drift caused by fine-tuning and greatly reducing trainable parameters.
50%UnverifiedAcross evaluations with five random seeds, APRIL led the compared continual learning methods on final-session average scores.
50%UnverifiedAPRIL's final-session average scores on the ICBHI dataset were 64.4% for forward and 65.4% for reverse.
50%UnverifiedAPRIL's final-session average scores on the SPRSound dataset were 83.1% for forward and 82.2% for reverse.
50%UnverifiedUnder all tested protocols, APRIL's average forgetting rate was below 1.0%.
50%UnverifiedAPRIL's replay-free nature aligns with the privacy requirements of medical scenarios, providing a feasible path for deploying continuous breath sound classification in real medical environments.
50%