Research map / Architectures and efficiency / CNN design
msf-CNN: Patch-based Multi-Stage Fusion with Convolutional Neural Networks for TinyML
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 msf-CNN, a DAG-based shortest-path search over multi-stage patch-based fusion settings for CNNs on microcontrollers, cutting peak RAM by up to ~50% or more versus MCUNetV2 and StreamNet.
Problem
Reducing peak RAM and balancing inference latency for CNN inference on microcontrollers, where existing patch-based fusion work considers only a single fusion block and limited solution spaces.
Approach
Model the CNN as an inverted dataflow DAG whose edges are single layers or candidate fusion blocks annotated with RAM and MAC costs, then find optimal fusion settings via minimax/shortest-path algorithms with iterative edge pruning under RAM or compute constraints; also rewrite global pooling and dense layers iteratively to cut RAM.
Key results
- On MCU boards, msf-CNN reduces peak RAM for MBV2-w0.35 to 8.56 kB versus 63 kB (MCUNetV2) and 66 kB (StreamNet), a 65-87% reduction.
- Under P1 with compute limit 1.1, msf-CNN reaches 67.9 kB RAM for MBV2-w0.35 vs 194.44 kB vanilla; unconstrained fusion gives 1.6x-2.7x compute overhead.
- msf-CNN keeps Top-1 accuracy identical to vanilla models (e.g., 48.94% for MBV2-w0.35) because it only changes scheduling, and latency rises 2x-5x versus vanilla at minimal RAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use msf-CNN fusion with minimal peak-RAM settings instead of the strongest listed fusion baseline | MBV2-w0.35 · MBV2-w0.35 | peak RAM (kB): 63 -> 8.56 | Helped |
| Use msf-CNN fusion with minimal peak-RAM settings instead of the strongest listed fusion baseline | MN2-vww5 · MN2-vww5 | peak RAM (kB): 44 -> 15.368 | Helped |
| Use msf-CNN fusion with minimal peak-RAM settings instead of the strongest listed fusion baseline | MN2-320K · MN2-320K | peak RAM (kB): 208 -> 51.164 | Helped |
| Use minimal-RAM fusion settings rather than vanilla inference | msf-CNN · MBV2-w0.35, MN2-vww5, MN2-320K | latency: $2 \times$ to $5 \times$ | Hurt |
| Apply msf-CNN fusion rather than vanilla execution | MBV2-w0.35 · MBV2-w0.35 | Top-1 accuracy: 48.94~\% -> 48.94~\% | No clear effect |
| Apply msf-CNN fusion rather than vanilla execution | MN2-vww5 · MN2-vww5 | Top-1 accuracy: 88.89~\% -> 88.89~\% | No clear effect |
| Apply msf-CNN fusion rather than vanilla execution | MN2-320K · MN2-320K | Top-1 accuracy: 61.76~\% -> 61.76~\% | No clear effect |
| Replace standard global pooling with iterative global pooling (ablation) | msf-CNN global pooling · 7 × 7 global pooling | RAM usage: 2\% of the original size | Helped |
| Replace standard dense computation with iterative dense computation (ablation) | msf-CNN dense layer · 1024→256 dense layer | RAM usage: 20\% of the original | Helped |
Surprising findings
- Minimal-RAM configurations increase inference latency by 2x to 5x over vanilla, making them suitable only for latency-tolerant applications.
- msf-CNN underperforms StreamNet on latency because it uses a 1-D cache and no hardware-specific (CMSIS) acceleration.
Limitations
- Only H-cache scheme is supported; 2-D cache and alternative caching strategies are not explored.
- Only convolutional architectures are covered; RNNs and transformers are not supported in the current implementation.
- Latency is higher than StreamNet because the implementation lacks hardware-specific libraries and acceleration instructions.
Experimental setup
Datasets: ImageNet Visual Wake Words (VWW)
Benchmarks: ImageNet Visual Wake Words (VWW)
Metrics: peak RAM (kB) compute overhead factor F inference latency (ms) Top-1 accuracy
Baselines: MCUNetV2 StreamNet vanilla un-fused models heuristic fusion of head layers
Builds on: MCUNetV2 StreamNet microTVM RIOT-ML Dijkstra's algorithm
Closest papers by meaning
- DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators (NeurIPS 2024 · similarity 0.85)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · similarity 0.83)
- FFN Fusion: Rethinking Sequential Computation in Large Language Models (NeurIPS 2025 · similarity 0.83)
- SMMF: Square-Matricized Momentum Factorization for Memory-Efficient Optimization (AAAI 2025 · similarity 0.83)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.83)
- M*: A Modular, Extensible, Serving System for Multimodal Models (NeurIPS 2026 · similarity 0.82)
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.