Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

SenseNet: Neural Architecture Search Inspired by Adaptive Biological Sensing for Transparency and Adaptability

Neha Bhargava¹, Kirtana Sunil Phatnani²; ¹Motilal Oswal Financial Services Ltd, ²Fractal Analytics

Presenter: Kirtana Sunil Phatnani

Neural Architecture Search (NAS) has emerged as a key method for automating neural network design. However, most existing approaches rely on static optimization strategies that struggle to adapt to new tasks, incorporate real-time feedback, or explain their decision processes—limitations that hinder performance in dynamic environments. Inspired by evolutionary phenomena and human learning — where organisms develop or adapt sensory systems to interpret and act on environmental cues—this work explores how NAS can incorporate similar mechanisms to improve adaptability and dynamic search strategy. This paper presents a hybrid approach called SenseNet that combines large language model (LLM)-driven explainability, dynamic optimization, and real-time adaptability to enhance NAS decision-making. At its core, SenseNet features a meta-controller that dynamically selects high-level strategy - exploration, exploitation, or balanced — based on sensing environmental cues, while an ML strategist translates these high-level strategies into tailored crossover and mutation operations that guide architecture evolution effectively. We evaluate SenseNet comprehensively on NATS-Bench and TransNAS-Bench, demonstrating its adaptability and effectiveness. Our experiments show that SenseNet achieves state-of-the-art results on NATS-Bench and performs competitively on TransNAS-Bench. By embedding sensing and response mechanisms into NAS, SenseNet enhances both efficiency and transparency in neural network optimization, shifting the paradigm from rigid search techniques to biologically inspired, self-adaptive, and transparent neural network optimization.

Topic Area: Methods & Computational Tools

Extended Abstract: Full Text PDF