21.09.2017
Burcu Kaptıkaçtı
Index
 What is DeepLearning?
 History
 What changed?
 Where we use?
 DeepLeaarning Architectures
 DeepLearning Libraries
 Demo
What is DeepLearning?
 A powerful class of machine learning model
 Multi-layered Neural Network
What is DeepLearning?
History
<<Hype Cycle>>
What Changed?
What Changed?
 Amount of data
What Changed?
 Amount of data
 GPU
What Changed?
 Amount of Data
 GPU
 Multi-layered Structure
Where we use?
 Speech Recognition
 LSTM
 Cortana,Xbox,Skype Translator, Google Now, Siri, Baidu
 Image Recognition
 CNN
 ImageNet,MNIST
 Natural Language Processing
 Word Embedding
Where we use?
 https://www.youtube.com/watch?v=MVBe6_o4cMI
 https://www.youtube.com/watch?v=Voz4dosVGSM
 https://www.youtube.com/watch?v=V1eYniJ0Rnk
Neural Networks
Neural Networks
Neural Networks
DeepLearning Architectures
 Convolutional Neural Networks (CNN)
 Recurrent Neural Networks (RNN)
 Long Short Term Memort (LSTM)
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
 Kernel(K) or Filter
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN)
DeepLearning Libraries
 Caffe
DeepLearning Libraries
 Caffe
 Torch
DeepLearning Libraries
 Caffe
 Torch
 Theano
Deep Learning Libraries
 Caffe
 Torch
 Theano
 TensorFlow
DeepLearning Libraries
 Caffe
 Torch
 Theano
 TensorFlow
 NDIVA DIGIT
DeepLearning Libraries
 Caffe
 Torch
 Theano
 TensorFlow
 NDIVA DIGIT
 Keras
 http://deeplearning.net/software_links/
DeepLearning Libraries
 Caffe2 --Facebook
 Torch --Facebook
 TensorFlow --Google
 CNTK--Microsoft
 MxNet– Amazon/Apache
Deep Learning Libraries
Knet
 Koç University deep learning framework
 https://github.com/denizyuret/Knet.jl
Keras
 Python Deep Learning Library
 Documentation
 https://keras.io/
Keras
 Sequential Model
 Functional Model
Keras-Sequential Model
Keras-Sequential Model
 Input
 Output
Keras-Sequential Model
 Compile
model.compile(optimizer=... , loss=..., metrics=[...])
Keras- Optimization Functions
 'sgd': SGD
 'rmsprop': RMSprop
 'adagrad': Adagrad
 'adadelta': Adadelta
 'adam': Adam
 'adamax': Adamax
 'nadam': Nadam
Loss Functions
 MSE
 Mean Absolute
 hinge
 squared_hinge
….
 https://keras.io/losses/
Keras –Loss Functions
 mean_squared_error / mse
 mean_absolute_error / mae
 mean_absolute_percentage_error / mape
 mean_squared_logarithmic_error / msle
 squared_hinge
 Hinge
model.compile(loss='categorical_crossentropy',...)
Activation Functions
 Sigmoid
 Softmax
 Tanh
 Relu
 https://keras.io/activations/
Keras - Save
model.fit(giris,cikis,epochs=1,validation_split=0.1)
model.save("kayıt")
model= keras.models.load_model("kayıt")
Overfitting
 Dropout
Keras
 Python Deep Learning Library
 Documentation
 https://keras.io/
 https://repo.continuum.io/archive/index.html
 conda install -c conda-forge keras
Datasets
 https://archive.ics.uci.edu/ml/datasets/
 https://datamarket.com/data/
 https://www.cs.toronto.edu/~kriz/cifar.html
Demo -1
 Dataset
 https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+(Diagnostic)

Deep Learning

Editor's Notes