Unified DNAS for Compressible (UDC) NNs
AI model · Computer vision · Image → Text by Tenstorrent
Unified DNAS for Compressible TinyML Models for Neural Processing Units
A neural architecture search framework for creating compressible TinyML models. It enables high-accuracy computer vision on memory-constrained IoT hardware by exploiting weight quantization and sparsity.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Launched
- 2023-02-10
Category: Computer Vision
- Tasks
- Image classification (ImageNet)
- Released
- 2023-02-10
- Target Hardware
- Neural Processing Units (NPU), IoT Hardware
- Optimization Methods
- Weight Quantization, Sparsity
- Benchmark Dataset
- ImageNet
- Model Size Reduction
- Up to 3.35x (iso-accuracy)
- Up to 3.35x smaller models at iso-accuracy
- 6.25% more accurate at iso-model size on ImageNet
- Optimized for low-cost IoT hardware and NPUs
- Exploits weight quantization and sparsity
Overview
Unified DNAS for Compressible (UDC) NNs is a neural architecture search framework developed by Tenstorrent to create optimized TinyML models. It is designed for developers and businesses deploying AI on low-cost IoT hardware and Neural Processing Units (NPUs) where device memory capacity is strictly limited. The framework addresses these hardware constraints by exploring a large design space to generate compressible neural networks that balance trade-offs between size and accuracy. UDC specifically exploits weight quantization and sparsity to fit a higher number of parameters into a smaller footprint. Benchmarks on ImageNet demonstrate that UDC-generated networks are up to 3.35x smaller at iso-accuracy or 6.25% more accurate at iso-model size compared to previous state-of-the-art work. This technology serves as a foundation for Tenstorrent's product development in efficient edge and embedded AI.
TinyML Optimization
- DNAS Design Space Exploration: The framework explores a large design space to generate state-of-the-art compressible neural networks specifically for Neural Processing Units.
- 3.35x Model Size Reduction: UDC networks achieve up to 3.35x smaller footprints at iso-accuracy compared to previous TinyML work.
- 6.25% Accuracy Improvement: The technology delivers up to 6.25% more accurate results at the same model size as existing solutions on ImageNet benchmarks.
Memory Efficiency Techniques
- Weight Quantization Support: UDC exploits weight quantization to fit more parameters into the same memory footprint of low-cost IoT hardware.
- Sparsity Exploitation: The framework utilizes sparsity to address the memory capacity challenges of limited device environments.
- IoT Hardware Compatibility: UDC is designed to deploy TinyML models on low-cost IoT hardware with limited device memory capacity.
Benchmarks
- ImageNet: up to 3.35x smaller (iso-accuracy) or 6.25% more accurate (iso-model size)
Use cases
- IoT developers deploy TinyML models on low-cost hardware with limited memory capacity using UDC-optimized networks.
- Hardware engineers use UDC to fit more complex neural network parameters into the same NPU footprint through weight quantization.
- Computer vision researchers achieve higher accuracy on ImageNet benchmarks without increasing the model size for edge devices.
Specifications
- Tasks
- Image classification (ImageNet)
- Released
- 2023-02-10
- Target Hardware
- Neural Processing Units (NPU), IoT Hardware
- Optimization Methods
- Weight Quantization, Sparsity
- Benchmark Dataset
- ImageNet
- Model Size Reduction
- Up to 3.35x (iso-accuracy)
- Accuracy Gain
- Up to 6.25% (iso-model size)
FAQ
What hardware is UDC designed for?
UDC is designed for low-cost IoT hardware and Neural Processing Units (NPUs) that have limited device memory capacity.
How does UDC handle memory constraints?
It uses model compression techniques, specifically weight quantization and sparsity, to fit more parameters into a smaller memory footprint.
What performance gains does UDC offer?
On ImageNet benchmarks, UDC networks are up to 3.35x smaller at the same accuracy level or 6.25% more accurate at the same model size compared to previous work.
What does UDC stand for?
UDC stands for Unified DNAS for Compressible neural networks, which is a framework for generating optimized TinyML models.
Added by AntSpace from public sources (tenstorrent.com).