M
Matthijs Hollemans
machinethink.net · Personal blog · on HN since 2017
Matthijs Hollemans writes technical articles on machine learning on Apple devices, including Core ML, neural networks, and mobile performance.
What HN saysHN commenters cite his benchmarks, such as MobileNet V2 speed on iPhone, and his Core ML training post as references in on-device ML discussions.
15HN people shared it
3linked it in comments
14stories submitted
516story points
5 years agolast shared
2017people per year2026
Most-shared posts
- 1One-shot object detection3 people · 5 points · first shared 2018 · discussion on HN
- 2The lost art of 3D rendering without shaders2 people · 323 points · first shared 2017 · 78 comments on HN
- 3A peek inside Core ML2 people · 4 points · first shared 2017 · discussion on HN
- 4YOLO: Core ML versus MPSNNGraph2 people · 2 points · first shared 2017 · discussion on HN
- 5Apple machine learning in 2020: What’s new?1 people · 95 points · first shared 2020 · 25 comments on HN
- 6Machine learning on mobile: on the device or in the cloud?1 people · 81 points · first shared 2017 · 19 comments on HN
- 7Convolutional neural networks on the iPhone with VGGNet1 people · 3 points · first shared 2017 · discussion on HN
- 8Compressing deep neural nets1 people · 2 points · first shared 2017 · discussion on HN
- 9New mobile neural network architectures benchmarked in iOS1 people · 1 points · first shared 2020 · discussion on HN
- 10Mobile architectures1 people · first shared 2021
- 11Coreml training part11 people · first shared 2020
- 12Mobilenet v21 people · first shared 2019
What HN people wrote when they linked it
nl · 2021-07-16
It's not misleading assuming you know the field.
If you don't you might not realize they are comparing ResNet (designed for ultimate performance) vs EffcientNet to show how close it gets in accuracy in the FLOPs budget.
Note that the best accuracy for EffcientNet is roughly 1-2% better than ResNet/ResNeXT/SENet/etc but does have a much better FLOPs budget.
But these other architectures were never optimised for FLOPs. It wasn't event a consideration when designing them.
And EffcientNet is about a (manually designed) technique for scaling neural networks up in accuracy. Only EffcientNet-B0 is designed by AutoML, the others are scaled up. See the paper[1] for complete details.
Like-for-like should be against MobileNet etc. EfficentNet is better, but the comparison is more reasonable in general.
https://machinethink.net/blog/mobile-architectures/ is a really good overview.
[1] https://arxiv.org/pdf/1905.11946.pdf
zitterbewegung · 2020-06-22
You can do some refinement right now already (from WWDC 2019).
But yes, most models are trained on NVIDIA GPUs that are deep learning. All of the other model types can already be trained on regular CPUs.
https://machinethink.net/blog/coreml-training-part1/
ithkuil · 2019-03-11
Is it?
The edge TPU can do MobileNet V2 at 100 FPS.
An iPhone 7 can do it at 145 FPS (source https://machinethink.net/blog/mobilenet-v2/)
The deal is though that the Edge TPU is able to do it at much lower power.