Scholarlink

This work presents a hardware-efficient framework for automated bird call identification, optimized for tinyML deployment. The pipeline uses multi-objective Bayesian optimization to design Pareto-optimal residual networks under a 2MB constraint. These compact student models are then enhanced via knowledge distillation from a high-capacity teacher, achieving an average 3.5% improvement in accuracy on BirdCLEF 2021. Final post-training quantization to int8 enables real-time on-device inference.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing TinyML Models for Bird Call Recognition via Multi-objective Bayesian Search and Knowledge Distillation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Agnivo Ghosh,
  • Soumen Garai,
  • Suman Samui

摘要

This work presents a hardware-efficient framework for automated bird call identification, optimized for tinyML deployment. The pipeline uses multi-objective Bayesian optimization to design Pareto-optimal residual networks under a 2MB constraint. These compact student models are then enhanced via knowledge distillation from a high-capacity teacher, achieving an average 3.5% improvement in accuracy on BirdCLEF 2021. Final post-training quantization to int8 enables real-time on-device inference.