Seq103 réduit les modèles de séquence 160 000 fois tout en conservant une précision de 82%

Up to 160,601x fewer parameters while holding onto 81.95% of best-baseline accuracy — that’s the headline number from Seq103, a new neuroevolution framework for sequence classifiers. On step-wise recurrent tasks, the parameter savings run 34.6x to 3,218x, with average accuracy retention at 86.96%. These aren’t cherry-picked results; they come from 8 text classification datasets plus the full UCRArchive2018 benchmark, all 128 univariate time-series datasets.

What Makes Seq103 Different

Seq103 is a NEAT-style neuroevolution approach — it evolves both the network topology and the weights using evolutionary algorithms. The clever bit is a shared evolutionary backbone that works for both step-wise recurrent and sample-wise feedforward tasks. When you need temporal memory (recurrent), the framework flips on an optional hidden-state extension with hidden nodes and hidden connections. For feedforward tasks, that extension is disabled, but the core search pipeline — elementary node-and-connection representation, per-class RMSE-based evaluation, mutation with class-wise recombination, and elitism — stays exactly the same.

Parameter Efficiency That Actually Means Something

On sample-wise tasks over UCRArchive2018, Seq103 retains 81.95% of the best-baseline accuracy while its models use between 11.8x and 160,601x fewer parameters. For step-wise tasks, the savings are 34.6x to 3,218x with 86.96% accuracy retention. Those ratios are not typos. A 160,601x compression means a baseline model with 16 million parameters gets replaced by a network of 100 parameters — for a 2% accuracy hit on average across 128 datasets. That level of compression makes deployment on edge devices and microcontrollers a real possibility.

What This Enables Next

Seq103 doesn’t just shrink models; it automates the architecture search for sequence classification without requiring a separate supernet or weight-sharing tricks. By unifying the search across recurrent and feedforward modes, it opens the door to tiny, task-specific sequence models that can run on hardware nobody would consider for a Transformer. Expect to see this kind of parameter efficiency applied to sensor data, wearables, and low-power embedded systems where every byte counts.

Source: Seq103: A Unified Neuroevolution Framework for Compact Sequence Architecture Discovery
Domain: arxiv.org