They're Good Dogs A gentle introduction to Core ML and Vision Andrew Harvey (@mootpointer)
A Question For You
What we will not cover • In depth machine learning • Cats • Bleeding edge real - world implementations
What we will cover • Dogs • The basics of the tools available • More dogs • What you could do with this stu ff
Pro Tip
"The best part of WWDC is the labs."
"The best part of WWDC is the labs ! "
Enough dogs. Show me the code (Said no one ever.)
xkcd.com/1425
Let's break it down
Your App
Domain Specific
Vision
Vision With a pixel buffer VNImageRequestHandler(cvPixelBuffer: imageBuffer, options: [:])
Vision Or a CGImage VNImageRequestHandler(cgImage: image, options: [:])
Natural language processing
GameplayKit
Core ML
Core ML The fun stuff! model.prediction(from: featureProvider)
MLFeatureProvider
Metal Performance Shaders
Accelerate and BNNs
If you're starting out You probably want Core ML
Any Questions?
That's enough code. More dogs!
Let's talk about Adversarial Examples
Here's one I prepared earlier
(Live demo ensues)
Ecosystem
You still have to train your own model
Keras keras.io
Caffe caffe.berkeleyvision.org
scikit-learn scikit-learn.org
coremltools developer.apple.com/machine-learning
So what?
Privacy
Latency
Offline
"If you can avoid a network round trip, you probably should."
Core ML + ARKit = !
Where to from here?
Learn you some acronyms CNNs, LSTMs, RNNs. (Oh my!)
xkcd.com/1838
fast.ai
Make your own. developer.apple.com/machine-learning
Understand your features
Do No Evil
Make fun things!
Thanks! (I'm hiring!) andrew@canceraid.com
Recommend
More recommend