Research map / Architectures and efficiency / Quantization
UDC: Unified DNAS for Compressible TinyML Models for Neural Processing Units
Igor Fedorov, Ramon Matas, Hokchhay Tann, Chuteng Zhou, Matthew Mattina, Paul Whatmough
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 |
|---|---|---|---|
| Increase Gumbel-softmax temperature tau from 0.66 to 10 (ablation) | Vanilla DNAS · MBNetV2 search space | L_E^z-hat: 0.04 -> 0.53 | Hurt |
| Add projection onto exploration set (xi=0.5) to DNAS (ablation) | Vanilla DNAS · MBNetV2, tau=0.66 | L_E^z-hat: 0.04 -> 0.33 | Hurt |
| Add rejection sampling (theta=0.99) on top of projection (ablation) | UDC · MBNetV2, tau=0.66 | L_E^z-hat: 0.33 -> 0.18 | Improved |
| Add rejection sampling (theta=0.5) on top of projection (ablation) | UDC · MBNetV2, tau=0.66 | L_E^z-hat: 0.33 -> 0.27 | Improved |
| Use rejection sampling at high temperature tau=10 (theta=0.99) (ablation) | UDC · MBNetV2, tau=10 | L_E^z-hat: 0.61 -> 0.59 | No clear effect |
| Use 32 MC samples instead of 1 with fewer steps (ablation) | UDC · DNAS search speed | speed-up: 5.3x speed-up | Improved |
| Disable UDC components (partial UDC variant) (ablation) | UDC · CIFAR100 0.5MB | accuracy (%): 79.24 -> 76.48 | Hurt |
| Disable a UDC component requiring lambda tuning (partial UDC) (ablation) | UDC · CIFAR100 0.5MB | GPUD: 1.17 -> 6.21 | Hurt |
| Use vanilla DNAS instead of UDC (ablation) | Vanilla DNAS · CIFAR100/ImageNet | HW constraint satisfaction | Hurt |
| Replace uniform quantization with proposed number format that quantizes beyond pruning boundary (ablation) | UDC · ImageNet | weight norm growth: 4.6x -> 2.2x | Improved |
| Use proposed number format instead of uniform quantization (ablation) | UDC · ImageNet | accuracy | Improved |
| Deploy with b*=8 uniform quantization of trained model | UDC · ImageNet | accuracy | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MCUNet — MCUNet: Tiny Deep Learning on IoT Devices (NeurIPS 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- HALOC: Hardware-Aware Automatic Low-Rank Compression for Compact Neural Networks (AAAI 2023 · similarity 0.84)
- "Lossless" Compression of Deep Neural Networks: A High-dimensional Neural Tangent Kernel Approach (NeurIPS 2022 · similarity 0.84)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · similarity 0.83)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.82)
- Analog Foundation Models (NeurIPS 2025 · similarity 0.82)
- TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices (NeurIPS 2024 · 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.
