TCUQ: Single-Pass Uncertainty Quantification from Temporal Consistency with Streaming Conformal Calibration for TinyML

2025-08-18

AI Summary

Uncertainty quantification for TinyML devices remains challenging under resource constraints, streaming inputs, and unlabeled data. Method: We propose the first one-pass, unsupervised uncertainty monitoring framework that jointly leverages short-term temporal consistency and streaming conformal prediction. It constructs a dynamic risk score from model outputs and lightweight signal features, and employs a circular buffer (O(W) memory) with O(1) incremental updates in a streaming conformal layer to enable calibrated abstention decisions. Contribution/Results: This work is the first to integrate temporal consistency into streaming conformal inference, achieving calibration, ultra-low latency, and minimal resource overhead. On microcontrollers, it reduces memory footprint by 50–60% and accelerates inference by 30–45%. Out-of-distribution detection improves by 3–7 percentage points in AUPRC; high-risk scenario AUPRC reaches 0.86, and failure detection AUROC achieves 0.92.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Data Mining & Knowledge Management: Anomaly/Outlier Detection

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

We introduce TCUQ, a single pass, label free uncertainty monitor for streaming TinyML that converts short horizon temporal consistency captured via lightweight signals on posteriors and features into a calibrated risk score with an O(W ) ring buffer and O(1) per step updates. A streaming conformal layer turns this score into a budgeted accept/abstain rule, yielding calibrated behavior without online labels or extra forward passes. On microcontrollers, TCUQ fits comfortably on kilobyte scale devices and reduces footprint and latency versus early exit and deep ensembles (typically about 50 to 60% smaller and about 30 to 45% faster), while methods of similar accuracy often run out of memory. Under corrupted in distribution streams, TCUQ improves accuracy drop detection by 3 to 7 AUPRC points and reaches up to 0.86 AUPRC at high severities; for failure detection it attains up to 0.92 AUROC. These results show that temporal consistency, coupled with streaming conformal calibration, provides a practical and resource efficient foundation for on device monitoring in TinyML.

Problem

• Develops a single-pass uncertainty monitor for TinyML streaming
• Converts temporal consistency into calibrated risk scores efficiently
• Improves accuracy drop detection in corrupted data streams

Innovation

• Single-pass uncertainty quantification via temporal consistency
• Streaming conformal calibration for label-free monitoring
• Lightweight O(W) buffer and O(1) updates for TinyML