- Date: June 7 2017
- Venue:
Tensorflow - Magnus Hyttsten and Marion Le Borgne
Google Developer Group Silicon Valley 320 Google Developers
Open source machine learning Magnus Hyttsten @magnushyttsten
What’s up today Thoughts on deep learning Open Source Models Image classification Natural language processing Generate your own artwork Resources you can use
https://www.youtube.com/watch?v=9ziVGkt8Gg4 Most of the slides are the same as the above video on google.io, Josh Gordon
No slide available online later.
I’m taking pictures. Slide 1
Slide 2
Slide 3: Cat or Dog
Slide 4: Tensorflow
Slide 5: Graph
Slide 6:
Slide 7:
Developing with Deep Learning
Slide 8:
Slide 9: Keras -> sits on top of TF TF-learn Code example
Slide 10: Slide 11:
goo.gl/TjQPfS https://github.com/random-forests/tensorflow-workshop/tree/master/extras
12: tensorboard, embedding
13: 3d tensorboard, clustering mnist
14: playground.tensorflow.org
15: TPU (tensor processing unit)
Search Jeff Dean 16: computer picture -> server room
17: server room 2
18: then switch to blog: mul and add cloud.google.com/blog/big-data/2017/05/an-in-depth-look-at-google-first-tensor-proessing-unit-tpu
19: inception 20: cloud.google.com/vision 21: using picture from new york 22: imagenet Dog breed and pictures 23: dog checking 24: graphi again Everything sums to 1
Transfer learning Throw the last layer out and use your 10 to 100 images and get it wo Replace output layer with nodes for your classes. Most of the network is unchanged Only last layer of weights are updated
25: application Tensorflow for poets.
26: design your experiment 27: 3 phones apps Blend artistic styles Classification Object detection 28: show and tell: description 29: deep dream 30. Style transfer Take the style + golden gate Paint golden gate bridge into picaso style 31.
- How does it work
Style, Content -> edges, textures, patterns, object, predictons, style loss, content loss, total loss
Target -> You can do this realtime Real dog camera video + sytle -> different
- Language SuntaxNet: Parsey McParseface, Parsey Saurus
I love NYC https://cloud/google.com/natural-language
Start a container - instructions at goo.gl/aH5wLP
“The gostak distims the doshes” Andrew Ingraham ->
From the internacctive text analyzer, insdie the syntaxnet container
Answer is verb Parse*tree, * = annotate_text(‘the gos…’) Token = parse_tree.token For t in tokens: If t.word == ‘distims’: Print ‘Answer is ‘, pos(t.tag)
Answer is: Verb
- What ifThe sentens is in german
load_model(‘data/German’)
All the languages are supported
More shared research in #tensorflow Neural audio synthesis,
Learning more Cs20si,stanford Cs231n.stanford
goo.gl/KewA03 -> machine learning recipes https://github.com/machine-learning-projects/machine-learning-recipes https://www.youtube.com/playlist?list=PLOU2XLYxmsIIuiBfYad6rFYQU_jL2ryal
Chris olah’s blog Colah.github.io *
Many tutorials: tensorflow.org
Temporal Anomaly Detection in Streaming Data with LSTM networks
Marion Le Borgne [email protected]
Explosion of sensor data
Preventative Maintenance Iot Sensors: Adjust energy usage in connected building Traffic Patterns: Identify unusual patterns from
Types of Anomalies Spatial Anomalies (out of bounds) Temporal Anomalies (sequences matter)
Scalable Anomaly Detection Ability to scale individual detectors to lots of data streams Ability to detect anomalies without prior knowledge Train
Anomaly Detection Benchmark numenta.com/numenta-anomaly-benchmark
Anonmaly Detection with LSTMs
Link to the slides will be available online.
