Seq103: A Unified Neuroevolution Framework for Compact Sequence Architecture Discovery

Seq103 introduces a unified neuroevolution framework that automatically discovers compact neural network architectures for sequence tasks, achieving 81-87% of baseline accuracy while using 11-3,200x fewer parameters. The framework applies the same evolutionary search pipeline to both recurrent and feedforward sequence classification, offering significant efficiency gains for resource-constrained deployments.

Seq103 addresses a fundamental challenge in machine learning: reducing model complexity without sacrificing performance. Traditional neural architecture search (NAS) methods often produce bloated models requiring substantial computational resources. This framework leverages neuroevolution—a technique that evolves both network structure and weights simultaneously—to discover inherently compact architectures that maintain competitive accuracy. The innovation lies in its unified approach: a single evolutionary pipeline handles both temporal (recurrent) and non-temporal (feedforward) sequence tasks through an optional hidden-state extension, eliminating the need for separate search processes.

The research builds on NEAT (NeuroEvolution of Augmenting Topologies), a foundational technique that gradually complexifies networks. Seq103's contribution centers on strict parameter efficiency constraints and class-wise evaluation metrics that prioritize compact solutions. Testing across 128 time-series datasets from UCRArchive2018 and 8 text classification datasets demonstrates broad applicability.

For practitioners, these results matter significantly. Machine learning deployments increasingly demand edge compatibility—mobile devices, IoT sensors, and resource-limited environments cannot run billion-parameter models. Seq103's ability to retain 80%+ baseline accuracy while reducing parameters by orders of magnitude makes sophisticated sequence modeling viable in constrained settings. The framework's generalizability across both recurrent and feedforward tasks increases its practical utility across different problem domains. This efficiency-accuracy trade-off positions automated architecture discovery as a viable tool for practitioners balancing performance requirements against deployment constraints, potentially accelerating adoption of ML in edge computing and embedded systems.

• →Seq103 achieves 81-87% of baseline accuracy while using 11x to 3,200x fewer parameters across sequence tasks
• →A unified evolutionary framework handles both recurrent and feedforward architectures through optional hidden-state extensions
• →Tested on 128 time-series datasets and 8 text classification benchmarks, demonstrating broad applicability
• →Enables compact neural networks suitable for edge deployment and resource-constrained environments
• →NEAT-based approach automates architecture discovery without manual design or separate search pipelines