Research map / Architectures and efficiency / CNN design
Memory-Optimal Direct Convolutions for Maximizing Classification Accuracy in Embedded Applications
Albert Gural, Boris Murmann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use herringbone instead of naive convolution (ablation) | CNN conv layers · Activation memory | bytes: 760.0 B -> 434.5 B | Improved |
| use replace method instead of naive (ablation) | CNN conv layers · Activation memory | bytes: 760.0 B -> 467.5 B | Improved |
| use herringbone instead of replace (ablation) | CNN conv layers · Activation memory | bytes: 467.5 B -> 434.5 B | Improved |
| use single transpose instead of herringbone (ablation) | CNN conv layers · Activation memory | bytes: 434.5 B -> 437.5 B | Hurt |
| use single transpose instead of herringbone (ablation) | CNN conv layers · Inference time | seconds: 819s -> 457s | Improved |
| use naive or replace instead of herringbone (ablation) | CNN conv layers · Inference time | seconds: 819s -> 396s | Improved |
| peak stack usage with single transpose (ablation) | CNN on Arduino · SRAM budget | spare bytes: 2B spare | Hurt |
| halve activation memory for given budget (naive vs herringbone) (ablation) | CNN · Validation error vs memory | validation error: 2x error | Improved |
| replace to herringbone method (ablation) | CNN · Validation error vs memory | relative error: ≈10% | Improved |
| decrease activation memory 10x below 1KB (ablation) | CNN · Validation error | validation error: 10x | Hurt |
| 2KB memory-optimal CNN instead of 2KB Bonsai | Bonsai · MNIST | accuracy (%): 94.38% -> 99.15% | Improved |
| 2KB CNN instead of 2KB ProtoNN | ProtoNN · MNIST | accuracy (%): 93.25% -> 99.15% | Improved |
Closest papers by meaning
- Designing Extremely Memory-Efficient CNNs for On-device Vision and Audio Tasks (IJCV 2026 · similarity 0.88)
- Resource-Efficient Neural Networks for Embedded Systems (JMLR 2024 · similarity 0.87)
- Constrained deep neural network architecture search for IoT devices accounting for hardware calibration (NeurIPS 2019 · similarity 0.86)
- SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Microcontrollers (NeurIPS 2019 · similarity 0.86)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.84)
- TinyFoA: Memory Efficient Forward-Only Algorithm for On-Device Learning (AAAI 2025 · similarity 0.83)
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.
