Making AI run really fast. Formerly @apple @berkeley_ai
Joined June 2013
- Tweets25
- Following18
- Followers377
- Likes324
Optimizing latency & accuracy for neural networks that fit on your phone! Introducing UPSCALE - a new export algorithm for pruned vision models with a 2.1% accuracy boost and up to 2X faster inference speeds! #ICML2023 🧵👇
Website: upscale.aaalv.in
ALT Presenting poster today 7/23 11a-12:30p at ICML Hall 1 #729!
4/5 - For latency, reordering is both beneficial and necessary. Without it, the pruned model is actually 🚨slower🚨 than the *original, unpruned model. Fortunately, reordering reduces latency by 8.6% on average, benefiting even tiny models such as SqueezeNet (-24.9%)📈
ALT Dashed line is the latency for a baseline unconstrained export, where you perform memory copies at inference time. Solid line is the latency for ours -- reordering channels in advance and slicing at inference time. We significantly reduce latency for the model across sparsity levels and heuristics. Depicted above is MobileNetV3-Small.
1
5
657
5/5 - Check out the original paper; if you’re at ICML, find me at Hall 1 #729 today 7/26, 11a-12:30p!
Paper: github.com/apple/ml-upscale
arXiv: arxiv.org/abs/2307.08771
Collaboration with: @HanxiangHao, @kaushikpatnaik, Sam Xu, @hadad_omer, @dgueraco, @jrenzhile, Qi Shan
7
693
Curious what explainable AI fails to explain? Check it out: youtu.be/fQ2eNFCSRiA or learn more about Neural-Backed Decision Trees on our website: nbdt.aaalv.in (just accepted by ICLR 2021)
Alvin Wan retweeted
Do we have to give up privacy to automatically track COVID-19 exposure? Several govt’s around the world have decided “yes,” but there’s an alternative. Just published 👉 medium.com/@covista/governme…
Alvin Wan retweeted
"Explaining what Explainable AI Did Not": @lvinwan describes how to make models as accurate as neural networks but with an interpretable decision process in the new BAIR blog!
bair.berkeley.edu/blog/2020/…
#xai
Just published in @TDataScience "What Explainable AI fails to explain (and how we fix that)" towardsdatascience.com/what-…. Feedback welcome!
Alvin Wan retweeted
In Part 1 of an upcoming article series by @lvinwan, you can learn how to build a simple, interactive VR model for an endless runner game — with webVR and a number of cool A-Frame concepts.
↬ Article: smashingmagazine.com/2019/03…
↬ Demo: ergo-1.glitch.me/
Alvin Wan retweeted
✏️ An introduction to programming in three dimensions and a guide to deploying a virtual reality model on the web, by @lvinwan:
smashingmagazine.com/2018/11…
I'm teaching a new online lesson "Build a Face-Swapping Filter Using Computer Vision" skl.sh/2PtwhBm via @skillshare!
Alvin Wan retweeted
✏️ What are the fundamentals of machine learning, and what are the necessary tools to evaluate risk and other concerns in a machine learning application? @lvinwan covers everything you need to get started: smashingmagazine.com/2018/09…
Alvin Wan retweeted
✏️ Building a desktop app that predicts which room you’re in just by using a simple machine learning algorithm: least squares. A tutorial by @lvinwan, here's how:
smashingmagazine.com/2018/08…
Alvin Wan retweeted
Setting up a LAMP server without frameworks: do.co/1Kk51sX tutorial by @lvinwan
For my grandpa and grandma. So that they'll never be forgotten, I put together an interactive website that'll... fb.me/1C3Q1rygV
3