Research map / Architectures and efficiency / RNNs and normalization
Position: AI Should Sense Better, Not Just Scale Bigger: Adaptive Sensing as a Paradigm Shift
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that AI should adapt sensor parameters at the input level (adaptive sensing) rather than relying solely on scaling models and datasets, with a roadmap, challenges, and a closed-loop MDP formulation for embodied agents.
Problem
Model-centric scaling of neural networks and datasets is costly (environmental, economic, equity) and fails under real-world covariate shifts from sensor and environment variation.
Approach
Treat sensor configuration (exposure, gain, modality weights) as a first-class, closed-loop adaptive input-level optimization that is guided by perception-aware quality metrics, so small models can match or beat much larger ones; extends single-shot Lens-style sensing to sequential sensorimotor and multimodal RL settings.
Key results
- Lens adaptive sensing improves accuracy by up to 47.58 percentage points without model modification (as reported from prior work).
- Preliminary evidence claims a 5M-parameter EfficientNet-B0 with ideal sensor adaptation can surpass the 632M-parameter OpenCLIP-H trained on 160x more data.
- Lens maintains high classification accuracy despite a 50x reduction in model size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive sensing with Lens instead of traditional human-oriented auto-exposure | Image classification · Image classification under sensor-induced covariate shifts | classification accuracy: up to 47.58%p | Helped |
| Use adaptive sensing with a model 50× smaller | Lens · Image classification | classification accuracy: 50$\times$ reduction in model size | Helped |
| Apply ideal sensor adaptation to EfficientNet-B0 rather than relying on the larger model alone | EfficientNet-B0 · Image classification | classification performance: 5M-parameter; 632M-parameter; 160$\times$ more | Helped |
| Coordinate RGB and depth sensing rather than using single-modal control or factory defaults | Adaptive cross-modal sensing · 6D object-pose estimation | accuracy | Helped |
| Use adaptive sensor parameters or viewpoints rather than non-adaptive configurations | Adaptive sensing · Object detection, segmentation, and rPPG under environmental shifts | robustness and accuracy | Helped |
Surprising findings
- High-quality images for model perception differ from those optimized for human perception.
- Adaptive sensing may enable a 5M-parameter model to surpass a 632M-parameter model trained on far more data.
Limitations
- No new experiments; empirical claims rely on prior work (Lens) and preliminary evidence.
- Lack of standardized benchmarks covering sensor and environmental variation beyond image classification.
- Privacy and ethical risks of real-time context-aware sensor control; integration barriers with existing pipelines.
Experimental setup
Datasets: ImageNet-ES ImageNet-ES-Diverse
Benchmarks: ImageNet-ES ImageNet-ES-Diverse
Metrics: classification accuracy
Baselines: OpenCLIP-H EfficientNet-B0 auto-exposure
Builds on: Lens Markov decision process EfficientNet
Closest papers by meaning
- Position: AI Scaling: From Up to Down and Out (ICML 2025 · similarity 0.83)
- AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.82)
- SEEA-R1: Tree-Structured Reinforcement Fine-Tuning for Self-Evolving Embodied Agents (NeurIPS 2025 · similarity 0.82)
- Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.81)
- Gradient Descent as Loss Landscape Navigation: a Normative Framework for Deriving Learning Rules (NeurIPS 2025 · similarity 0.81)
- ACTIVE-o3 : Empowering MLLMs with Active Perception via Pure Reinforcement Learning (ICML 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.