Research map / Architectures and efficiency / RNNs and normalization
TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices
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.
Proposes TinyTTA, a self-ensemble early-exit test-time adaptation framework with weight standardization and an MCU backpropagation library (TinyTTA Engine) that enables batch-size-1 TTA on microcontrollers and edge devices.
Problem
Test-time adaptation under distribution shift is memory-intensive (backpropagation through full networks), relies on normalization layers unsupported on MCUs, and degrades with the small batch sizes that constrained devices can afford.
Approach
Partition a pre-trained network into frozen submodules that approximate the full model, add early-exit heads with WS-normalized linear classifiers, and adapt only the exits using entropy-based early exiting so easy samples skip later computation; WS replaces batch-dependent normalization so adaptation works at batch size one.
Key results
- On STM32H747 (512 KB SRAM) with MCUNet on CIFAR10C, TinyTTA reaches 64.3% accuracy vs 60.2% inference-only, with latency 50.7 ms vs 55.8 ms and energy 11.5 mJ vs 12.7 mJ.
- On MobileNet/MCUNet MPU experiments, TinyTTA beats the second-best baseline by up to 57.6% accuracy (PACS, MCUNet) while using about 6x less memory than CoTTA on average.
- On Raspberry Pi Zero 2W with CIFAR10C, TinyTTA latency is 0.22 s/sample versus 0.25 s for EATA and 6.25 s for CoTTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply TinyTTA instead of no adaptation | MCUNet · CIFAR10C | accuracy: 60.2% -> 64.3% | Helped |
| apply TinyTTA instead of ECoTTA | MCUNet · CIFAR100C | accuracy: 46.6% | Helped |
| apply TinyTTA instead of ECoTTA | MCUNet · OfficeHome | accuracy: 52.1% | Helped |
| apply TinyTTA instead of ECoTTA | MCUNet · PACS | accuracy: 57.6% | Helped |
| apply TinyTTA instead of EATA | Raspberry Pi Zero 2W on CIFAR10C · CIFAR10C | latency: 0.25 -> 0.22 | Helped |
| apply TinyTTA instead of EATA | Raspberry Pi Zero 2W on CIFAR10C · CIFAR10C | energy consumption: 0.50 -> 0.44 | Helped |
| add on-device TinyTTA updates | MCUNet on STM32H747 · CIFAR10C | latency: 55.8ms -> 50.7ms | Helped |
| add on-device TinyTTA updates | MCUNet on STM32H747 · CIFAR10C | energy consumption: 12.7mJ -> 11.5mJ | Helped |
| replace WS normalization with GN (ablation) | TinyTTA on MCUNet · CIFAR10C L1 | accuracy: 83.2 -> 73.4 | Hurt |
| replace WS normalization with GN (ablation) | TinyTTA on MCUNet · CIFAR10C L2 | accuracy: 78.3 -> 67.2 | Hurt |
| replace WS normalization with GN (ablation) | TinyTTA on MCUNet · CIFAR10C L3 | accuracy: 73.6 -> 63.1 | Hurt |
| replace WS normalization with GN (ablation) | TinyTTA on MCUNet · CIFAR10C L4 | accuracy: 68.7 -> 57.9 | Hurt |
Surprising findings
- Modulation-based TTA methods (e.g., TENT) collapse to predicting few classes under severe corruption at batch size 1, and fine-tuning TTA collapses under mild shifts.
- Updating only BN/bias parameters (TinyTL-style bias-only TTA) is insufficient, performing much worse than updating early-exit heads.
Limitations
- Tested only with batch size one, on a single MCU type and image data (plus a short audio appendix).
- Requires retraining the model into the self-ensemble form before deployment.
- Entropy thresholds for early exits are dataset- and model-specific and require grid search.
Experimental setup
Datasets: CIFAR10C CIFAR100C OfficeHome PACS Musan Keywords Spotting Speech Commands V2
Benchmarks: CIFAR10C CIFAR100C OfficeHome PACS Musan Keywords Spotting
Metrics: accuracy memory usage (MB/SRAM/Flash) latency (s) energy (J)
Baselines: TENT CoTTA EATA EcoTTA Test-Time Training TinyEngine
Builds on: TENT EcoTTA BranchyNet TinyTL TensorFlow Lite Micro Weight Standardization
Closest papers by meaning
- SNAP: Low-Latency Test-Time Adaptation with Sparse Updates (NeurIPS 2025 · similarity 0.88)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.88)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · similarity 0.87)
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.85)
- DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators (NeurIPS 2024 · similarity 0.85)
- TinyFoA: Memory Efficient Forward-Only Algorithm for On-Device Learning (AAAI 2025 · similarity 0.84)
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.