ITP-NYU :: 4/7/2016
Convolutional neural networks
- [Nick Hubbard] Trolley problem (8:15)
- 'Animals are machines' - Rene Descartes (20:30)
- Wekinator custom outputs + Ableton (25:10)
- Wekinator + Audio Units (mac) (40:48)
- Wekinator + audioreactive Jitter (43:17)
- Kinect + gesture recognition options (52:38)
- Limitations of ordinary neural nets (1:03:16)
- Convolutional and pooling layers (1:24:51)
- Convnets: the whole pipeline (1:49:24)
Class notes
News / admin
- office hours
- official (M/T/W ?)
- unofficial at DBRSLabs
- i want to meet everyone
- alt-ai exhibition
Critical issues
- Nick (Trolley problem)
- Gene (Descartes on brutes)
- bias, ownership, privacy (1) (2) (3) (4)
- Critical algorithm studies
Advanced Wekinator examples + Gesture recognition with Kinect + openFrameworks
- Wekinator pt. 2
- Quick review
- inputs: Kinect, FaceOSC, bark/mfcc
- AbletonOSC
- AudioUnitOSC
- max transpose visual
- Kinect gesture recognition
- Pose detection (almost ready)
- Kinect2Gesture (see demo)
- ofxGRT
Introduction to convolutional neural networks
- Limitations of neural networks
- CIFAR-10 accuracy
- Q: how do normal neural nets deal with translation/stretch?
- How convnets work
- Convolution + pooling layers
- Classification demo
- Interpretation of n-1 layer
- Convnets searches Instagram!
Scientific applications of convnets(not in video)- Improved image classification
- -> localization, segmentation
- ensemble systems (annotation via RNN)
- Artistic applications (lecture 4)
- Visualization
- Deepdream, style transfer
- Visualization with t-SNE
- Transfer learning: Convnet -> Wekinator
Deep learning landscape (not in video)
- Main platforms
- Caffe, Theano, Torch, TensorFlow
- Wrappers :)
- Keras, tflearn
- Lasagne, nolearn
- openFrameworks support