A new evolutionary search method cuts the time to design efficient TinyML models by more than half, without needing an exact accuracy score for every candidate.
Researchers built TGL-NSGA-II, a search framework that swaps expensive full model evaluations for a faster ranking trick. A pretrained teacher model sorts training samples by difficulty and class, then each candidate architecture gets a short, capped round of knowledge distillation before being scored on a matching evaluation set. That teacher-guided score is blended with a Gaussian-process surrogate to decide which candidates earn a full evaluation. On keyword-spotting and bird-call classification tasks, the resulting rankings hit Kendall-tau correlations of 0.74 and 0.62, both above the method's own predicted lower bounds, and stratifying samples by difficulty cut proxy-score variance by 41% compared to random evaluation.
The insight here is that evolutionary search doesn't need to know exactly how good a candidate is, it just needs to know which candidate beats which. That's a cheaper problem to solve, and it's why TGL-NSGA-II beat full NSGA-II on hypervolume and generational distance for keyword spotting, posted the lowest false-positive rate on the BirdCLEF bird-call benchmark, and ran 2.2 times faster overall.
The method does depend on picking a well-matched teacher model: in a separate test, deliberately mismatching the teacher dragged rank accuracy down to a Kendall-tau of 0.41. But that's a tunable input, not a fundamental limit. For teams working within TinyML's punishing compute budgets, a 2.2x speedup with rankings this reliable is a real win, not just a shortcut that trades accuracy for speed.