2nd dl meetup @ dublin - irene
TRANSCRIPT
TensorFlow & Deep Learning
Tensorflow& Deep Learning
Irene LiPhD Student @ DITRA @ Barcelona Supercomputer Centre
OutlineDeep Learning ApplicationsTensorFlow- Introduction- Benefit- DemoMaterials
DL topicsImage recognitionSpeech recognitionVideo captioningNLP word embedding, TranslationBiological modelNeural ArtsGamesSelf-Driving CarsDeep Residual LearningGPUs
Neural ArtA Neural Algorithm of Artistic Style Leon A. Gatys, Alexander S. Ecker, Matthias Bethge
Neural Art
CNN for cv
VGG Model
Neural Art Style generationCode by Mark Chang, National Taiwan University https://github.com/ckmarkoh
Neural Art Style generationCode by Mark Chang, National Taiwan University https://github.com/ckmarkoh
Games Deep Q-Network (DQN)https://deepmind.com/dqn.htmlDeep Neural Network + Reinforcement Learninghow the game agents should act in an environment in order to maximize future cumulative reward (e.g., a game score)
These deep neural networks are trained by a novel combination of supervised learning from human expert games, and reinforcement learning from games of self-play.
Games Deep Q-Network (DQN)Naveen Appiah, Sagar Vare, Stanford.Playing FlappyBird with Deep Reinforcement Learning
Develop a CNN model to learn features from snapshots of the game and train the agent to take the right actions at each game instance.
Games Deep Q-Network (DQN)Naveen Appiah, Sagar Vare, Stanford.Playing FlappyBird with Deep Reinforcement Learning
Deep neural network to learn game specific features from raw pixels and decide on what actions to take.Instead, a reinforcement learning setting tries to evaluate the actions at a given state based on the reward it observes by executing it.
Reinforcement learning:Robots
Self-driving cars: NVIDIA DRIVE PX
DualNVIDIA Tegra X1 processorsdelivering a combined 2.3 TeraflopsInterfaces for up to 12 cameras, radar, lidar, and ultrasonic sensorsRich middleware for graphics, computer vision, and deep learning
This allows algorithms to accurately understand the full 360 degree environment around the car to produce a robust representation, including static and dynamic objects. Use of Deep Neural Networks (DNN) for the detection and classification of objects dramatically increases the accuracy of the resulting fused sensor data. - See more at: http://www.nvidia.com/object/drive-px.html#sthash.Vq3ZHH11.dpuf
Self-driving cars: NVIDIA DRIVE PX
This allows algorithms to accurately understand the full 360 degree environment around the car to produce a robust representation, including static and dynamic objects. Use of Deep Neural Networks (DNN) for the detection and classification of objects dramatically increases the accuracy of the resulting fused sensor data. - See more at: http://www.nvidia.com/object/drive-px.html#sthash.Vq3ZHH11.dpuf
Baidu Self Driving
Nvidia self driving
Object Detection
Under circustanes
Semantic Image Segmentation
Under circustanes
Racing cars: NVIDIA DRIVE PX2
150 Macbook Pro
Deep residual networkfor image recognitionMicrosoft Research:ImageNet computer vision challenge championship3.57% error on the ImageNet test set.152 layers of deep networks, 8 times deeper than VGGImage classification, detection and localization
Deep Residual Learning for Image Recognition http://arxiv.org/pdf/1512.03385v1.pdf
NVIDIA DGX-1:Worlds first DL supercomputer
Eight Tesla P100 GPU accelerators, 16GB memory per GPU
The combination of these software capabilities running on Pascal-powered Tesla GPUs allows applications to run 12x faster than with previous GPU-accelerated solutions.
TensorFlow2ed Generation of Distributed ML Systems
DL FrameworksTensorFlow: Google MXNET: dmlc Theano: LISA Torch: Facebook, DeepMindCNTK: Microsoft Caffe: Berkley, Google
TensorFlow:open source dL SystemCore in C++.Supports Python and C++.
Jeff Dean & Oriol Vinyals Google, NIPS, 2015
Computation:a Dataflow graphEdges: Flow of tensors.Graph of Nodes: Operations.
Jeff Dean & Oriol Vinyals Google, NIPS, 2015
DistributedMultiple Devices!Send and Receive Nodes
Jeff Dean & Oriol Vinyals Google, NIPS, 2015
Model Parallelism
Easy LifeEverything is in Tensor
Softmax Function
https://www.tensorflow.org/versions/r0.7/tutorials/mnist/beginners/index.html#mnist-for-ml-beginnersxbWMatmulAddSoftmax
Easy LifeGradient Descent:
Initialization & launch
https://www.tensorflow.org/versions/r0.7/tutorials/mnist/beginners/index.html#mnist-for-ml-beginners
A Few TensorFlow Community Examples DQN: github.com/nivwusquorum/tensorflow-deepq NeuralArt: github.com/woodrush/neural-art-tf Char RNN: github.com/sherjilozair/char-rnn-tensorflow Keras ported to TensorFlow: github.com/fchollet/keras Show and Tell: github.com/jazzsaxmafia/show_and_tell.tensorflow Mandarin translation: github.com/jikexueyuanwiki/tensorflow-zh
Useful linksWant to master DL?
NNDLSGDBPregularizationDropout
http://neuralnetworksanddeeplearning.com/
Deep Learning Book
http://deeplearningbook.org
Reinforcement LearningRichard SuttonAndrew Barto
online courseshttps://www.youtube.com/user/hugolarochelle (python)https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/ (torch)https://www.coursera.org/course/neuralnets (matlab)http://www0.cs.ucl.ac.uk/staff/d.silver/web/Teaching.html (rl)
TensorFlowGoogle open tensorflow in last yearhttp://tensorflow.org/https://github.com/tensorflow/tensorflowhttp://download.tensorflow.org/paper/whitepaper2015.pdf
Udacity course on tfhttps://www.udacity.com/course/deep-learning--ud730https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/udacityhttp://ireneli.eu/2016/03/13/tensorflow-04-implement-a-lenet-5-like-nn-to-classify-notmnist-images/
Referenceshttp://indico.cern.ch/event/510372/attachments/1245509/1840815/lecun-20160324-cern.pdfhttp://www.jianshu.com/notebooks/339523/latesthttp://people.idsia.ch/~juergen/blues/IDSIA-07-02.pdfhttp://deepdreamgenerator.com/https://github.com/phanein/deepwalkhttp://www.cs.toronto.edu/~rsalakhu/papers/annrev.pdfhttp://www.jianshu.com/p/8799afd9b7c7 http://www.jianshu.com/p/10d70c5ceb39