AI
Rank-Reliable Teacher-Guided Fitness Approximation for Expensive Evolutionary Optimization: A TinyML Architecture Search Study
Publication: cs.AI updates on arXiv.org
Author: Soumen Garai, Suman Samui
Source published: 28/09/2026, 04:00:00 UTC
Briefing created: 28/09/2026, 04:08:22 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/09/2026, 04:03:25 UTC.
Briefings
This study introduces Teacher-Guided Learning NSGA-II (TGL-NSGA-II), a low-fidelity TinyML architecture search framework that replaces exact fitness estimates with reliable ranking. By using a pretrained teacher to organize samples into difficulty-stratified groups, the method applies a capped knowledge-distillation procedure (KD-Lite) to candidates before scoring them on a separate evaluation set. This approach significantly reduces evaluation variance and improves ranking stability compared to standard methods, achieving superior hypervolume and speed on keyword spotting and bird-call classification tasks.
Key points
- TGL-NSGA-II uses a pretrained teacher to organize samples into strata defined by difficulty and class, then applies KD-Lite knowledge-distillation on a compact training set before scoring on a stratified evaluation set.
- On keyword spotting, the measured Kendall-$\tau$ value of 0.74 exceeds the predicted lower bound of 0.60, and on bird-call classification, it reaches 0.62 compared to a predicted lower bound of 0.46.
Context and limitations
The study focuses on population-level low-fidelity evaluation rather than convergence of the complete evolutionary trajectory. Results are limited to specific keyword spotting and bird-call classification tasks, and the capacity-adaptive distillation rule remains untested on other architectures.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/09/2026, 04:08:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibzbhcdpm7ohp5iyvdxldt7sbl4kzvinbj4vygkrcr27onrjlokaq