Research map / Architectures and efficiency / CNN design
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
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 DEX, which extends input channels of a CNN on tiny AI accelerators (MAX78000/78002) via patch-wise even sampling and channel-wise stacking, recovering more image information at no added latency.
Problem
Tiny AI accelerators with per-processor memory force downsampling of input images, causing information loss and accuracy degradation, while the input layer leaves most processors and data memory idle.
Approach
Use idle per-processor data memory and processors to stack evenly sampled pixels from each image patch across 64 input channels, so the first CNN layer sees much more of the original image at the same inference latency.
Key results
- DEX improves average accuracy by 3.5 %p over downsampling (62.0 vs 58.2 on ImageNette; 22.6 vs 19.4 on Caltech256) across four models and four datasets.
- DEX keeps on-accelerator inference latency unchanged (e.g., SimpleNet ~2591 us vs 2592 us) while raising model size by about 3.2% on average.
- DEX uses 21.3x more original image information and 100% processor utilization in the first layer versus 4.7% for downsampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| extend input data to 64 channels using patch-wise even sampling and channel-wise stacking | SimpleNet · ImageNette | accuracy: 57.8 ± 1.2 -> 61.4 ± 0.6 | Helped |
| extend input data to 64 channels using patch-wise even sampling and channel-wise stacking | SimpleNet · Caltech101 | accuracy: 54.6 ± 2.1 -> 56.9 ± 1.3 | Helped |
| extend input data to 64 channels using patch-wise even sampling and channel-wise stacking | SimpleNet · Caltech256 | accuracy: 19.8 ± 0.6 -> 22.8 ± 0.5 | Helped |
| extend input data to 64 channels using patch-wise even sampling and channel-wise stacking | SimpleNet · Food101 | accuracy: 16.0 ± 0.4 -> 18.4 ± 0.4 | Helped |
| add coordinate channels (CoordConv) | Downsampling baseline · four datasets, average accuracy | accuracy: 0.1%p degradation on average | Hurt |
| repeat the downsampled image across 64 channels (ablation) | DEX input extension · ImageNette | accuracy: 57.8 ± 1.2 -> 56.3 ± 0.8 | Hurt |
| rotate the downsampled image across 64 channels (ablation) | DEX input extension · ImageNette | accuracy: 56.3 ± 0.8 -> 55.4 ± 0.7 | Hurt |
| divide the original image into tiles and stack tiles across channels (ablation) | DEX input extension · ImageNette | accuracy: 57.8 ± 1.2 -> 39.4 ± 0.7 | Hurt |
| sample pixels sequentially within each patch instead of evenly (ablation) | DEX input extension · ImageNette | accuracy: 61.4 ± 0.6 -> 61.0 ± 1.5 | Hurt |
| sample pixels randomly within each patch instead of evenly (ablation) | DEX input extension · ImageNette | accuracy: 61.4 ± 0.6 -> 60.4 ± 1.0 | Hurt |
| increase the number of input channels from 3 to 64 with DEX (ablation) | CNN models on tiny AI accelerators · four vision datasets | inference latency | No clear effect |
| increase channel count from 3 to 64 (ablation) | SimpleNet · ImageNette | accuracy | Helped |
Surprising findings
- CoordConv does not improve over downsampling on these tasks (about 0.1 %p lower on average).
- Repetition of the same downsampled image across 64 channels gives no gain over downsampling.
Limitations
- Only the first CNN layer is modified; effectiveness on larger models is not evaluated.
- May be less effective for tasks where additional pixel information is not beneficial.
- Power consumption increases with channel count (e.g., SimpleNet 53.8 to 68.9 mW).
Experimental setup
Datasets: ImageNette Caltech101 Caltech256 Food101
Benchmarks: ImageNette Caltech101 Caltech256 Food101
Metrics: classification accuracy inference latency model size processor utilization information ratio
Baselines: Downsampling CoordConv CoordConv (r) Repetition Rotation Tile
Builds on: CoordConv AI8X training framework
Closest papers by meaning
- Memory-efficient Patch-based Inference for Tiny Deep Learning (NeurIPS 2021 · similarity 0.86)
- StreamNet: Memory-Efficient Streaming Tiny Deep Learning Inference on the Microcontroller (NeurIPS 2023 · similarity 0.86)
- msf-CNN: Patch-based Multi-Stage Fusion with Convolutional Neural Networks for TinyML (NeurIPS 2025 · similarity 0.85)
- TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices (NeurIPS 2024 · similarity 0.85)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.85)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (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.