Видео

once-for-all: ofa.mit.edu mcunet: mcunet.mit.edu tinyTL: tinyml.mit.edu/projects/tinyml/tinyTL/

Presentation at Duke University, NSF AI Institute for Edge Computing: This video covers three TinyML papers: 1. MCUNet: Tiny Deep Learning on IoT Devices [NeurIPS’20] 2. MCUNetV2: Memory-Efficient Patch-based Inference for Tiny Deep Learning [NeurIPS’21] 3. Network Augmentation for Tiny Deep Learning [ICLR’22] slides: www.dropbox.com/s/n1evxyszq70pftv/Duke Athena talk MCUNet.pdf?dl=0

MIT Prof Song Han on TinyML: Reducing AIs Carbon Footprint. Stanford MLSys Seminar Episode 9

MCUNet: computation-efficient inference on device [1] TinyTL: computation-efficient transfer learning on device [2] DiffAugment: data-efficient GAN training [3] [1] MCUNet: Tiny Deep Learning on IoT Devices, arxiv.org/pdf/2007.10319.pdf [2] TinyTL: Reduce Memory, Not Parameters for Efficient On-Device Learning, arxiv.org/pdf/2007.11622.pdf [3] Differentiable Augmentation for Data-Efficient GAN ...

Presentation at edge ai vision alliance: www.edge-ai-vision.com/2020/08/once-for-all-dnns-simplifying-design-of-efficient-models-for-diverse-hardware-a-presentation-from-mit/

Transformers are essential for NLP applications. However, they are challenging to be deployed on mobile devices due to the intensive computation and memory requirement. We solve the challenge from two aspects: hardware-aware neural architecture search (NAS) and new primitive design. On one hand, to enable low-latency inference on resource-constrained hardware platforms, we propose to design Har...

Once for All: Train One Network and Specialize it for Efficient Deployment, ICLR'2020 #NAS, #TinyML, #EfficientAI Website: ofa.mit.edu Paper: arxiv.org/pdf/1908.09791.pdf Code: github.com/mit-han-lab/once-for-all

Song Han received the Ph.D. degree from Stanford University advised by Prof. Bill Dally. His research focuses on energy-efficient deep learning, at the intersection between machine learning and computer architecture. Song will be joining MIT as assistant professor in July 2018.

FPGA'17 tutorial

Neural networks are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. Song Han explains how deep compression addresses this limitation by reducing the storage requirement of neural networks by 10x-49x without affecting their accuracy and proposes an energy-efficient inference engine (EIE) that accelerates on the spars...

Song Han presented on 43rd International Symposium on Computer Architecture (ISCA'16) isca2016.eecs.umich.edu

Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce “deep compression”, a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35× to 49× without affecting ...

Deep compression and EIE: Deep learning model compression, design space exploration and hardware acceleration.