Introduction to Deep Learning
CS468 Spring 2017
Charles Qi
What is Deep Learning?
Deep learning allows computational models that are
composed of multiple processing layers to learn
representations of data with multiple levels of abstraction.
Deep Learning by Y. LeCun et al. Nature 2015
Artificial Machine Deep
Intelligence Learning Learning
From Y. LeCun’s Slides
Image: HoG Audio: Spectrogram
Image: SIFT Point Cloud: PFH
Linear Regression
SVM
Decision Trees
Random Forest
...
From Y. LeCun’s Slides
Image Video 3D CAD Model
Thermal Infrared Depth Scan Audio
Can we automatically learn “good” feature representations?
From Y. LeCun’s Slides
From Y. LeCun’s Slides
From Y. LeCun’s Slides
From Y. LeCun’s Slides
ImageNet 1000 class image classification accuracy
Big Data + Representation Learning with Deep Nets
By Google DeepMind
Acoustic Modeling
Near human-level
Text-To-Speech performance
https://deepmind.com/blog/wavenet-generative-model-raw-audio/
Big Data + Representation Learning with Deep Nets
Neural Translation Machine
by Quac V. Le et al at Google Brain.
https://research.googleblog.com/2016/09/a-neural-network-for-machine.html
Outline
● Motivation
● A Simple Neural Network
● Ideas in Deep Net Architectures
● Ideas in Deep Net Optimization
● Practicals and Resources
Outline
● Motivation
● A Simple Neural Network
● Ideas in Deep Net Architectures
● Ideas in Deep Net Optimization
● Practicals and Resources
A Simple Neural Network
Use recent three days’ average temperature to predict tomorrow’s
average temperature.
Image from CS231N
Sigmoid function
A Simple Neural Network
W1, b1, W2, b2, W3, b3
are network parameters
that need to be learned.
From CS231N
Neural Network: Forward Pass
x(1): 73.8
x(2): 75.8 y’: -10.4
x(3): 78.2
From CS231N
Neural Network: Backward Pass
73.8
75.8 Prediction: -10.4
Ground truth: 80.8
78.2
L2 error = (80.8 - (-10.4))^2
Update Network Parameters
Minimize:
Given N training pairs:
Neural Network: Backward Pass
Minimize:
Given N training pairs:
Sigmoid function
Non-convex optimization :(
Neural Network: Backward Pass
Minimize:
Given N training pairs:
Non-convex optimization :(
Use gradient descent!
Parameter update example:
A Simple Neural Network
Model: Multi-Layer Perceptron (MLP)
Loss function: L2 loss
Optimization: Gradient descent
Outline
● Motivation
● A Simple Neural Network
● Ideas in Deep Net Architectures
● Ideas in Deep Net Optimization
● Practicals and Resources
What people think I am doing when I What I actually do...
“build a deep learning model”
Contents
Building blocks: fully connected, ReLU, conv, pooling, upconv,
dilated conv
Classic architectures: MLP, LeNet, AlexNet, NIN, VGG,
GoogleNet, ResNet, FCN
Multi-Layer Perceptron
http://playground.tensorflow.org/
Fully Connected
Non-linear Op
Fully Connected
● The first learning machine: the
Perceptron Built at Cornell in 1960
● The Perceptron was a (binary) linear
classifier on top of a simple feature
extractor From LeCun’s Slides
From CS231N
Non-linear Op
Tanh
Sigmoid
Major drawbacks: Sigmoids saturate and kill gradients
From CS231N
Non-linear Op
ReLU (Rectified Linear Unit)
A plot from Krizhevsky et al. paper
+ Cheaper (linear) compared with Sigmoids (exp) indicating the 6x improvement in
+ No gradient saturation, faster in convergence convergence with the ReLU unit
- “Dead” neurons if learning rate set too high compared to the tanh unit.
Other Non-linear Op:
Leaky ReLU,
MaxOut From CS231N
Convolutional Neural Network : LeNet (1998 by LeCun et al.) Fully Connected
Non-linear Op
Convolution
Pooling
(pooling)
One of the first successful
applications of CNN.
Convolution Shared Weights & Convolutions:
Exploiting Stationarity
Fully Connected NN in high dimension
Slide from LeCun
Convolution
Stride 1 Stride 2
From CS231N
Pad 1 Pad 1
Stride 2 Stride 1
From vdumoulin/conv_arithmetic
Convolution
5x5 RGB Image
5x5x3 array
3x3 kernel, 2 output
channels, pad 1, stride 2
weights: 2x3x3x3 array
Pad 1 bias: 2x1 array
Stride 2
Output
3x3x2 array
H’ = (H - K)/stride_H + 1
= (7-3)/2 + 1 = 3
From CS231N
Pooling
Pooling layer (usually inserted in between conv layers) is
used to reduce spatial size of the input, thus reduce
number of parameters and overfitting.
Discarding pooling layers has been found to be important in training good generative models,
such as variational autoencoders (VAEs) or generative adversarial networks (GANs).
It seems likely that future architectures will feature very few to no pooling layers.
From CS231N
LeNet (1998 by LeCun et al.) Fully Connected
Non-linear Op
Convolution
Pooling
(pooling)
AlexNet (2012 by Krizhevsky et al.)
The first work that
popularized
Convolutional Networks
in Computer Vision
What’s different?
AlexNet (2012 by Krizhevsky et al.)
What’s different? Our network takes between five and six days
to train on two GTX 580 3GB GPUs. -- Alex
● Big data: ImageNet
● GPU implementation: more than 10x speedup
● Algorithm improvement: deeper network, data
augmentation, ReLU, dropout, normalization layers etc.
Network in Network (2013 by Min Lin et al.)
56x56x128 256x5x5x128 weights 256x5x5x256 weights 256x5x5x256 weights
Network in Network (2013 by Min Lin et al.)
256x5x5x128 weights
+ 1x1 conv (256x256 weights)
56x56x128 256x5x5x128 weights + 1x1 conv (256x256 weights)
1x1 convolution: MLP in each pixel’s channels
Use very little parameters for large model capacity.
VGG (2014 by Simonyan and Zisserman) The runner-up in ILSVRC 2014
Karen Simonyan, Andrew Zisserman: Very Deep
Convolutional Networks for Large-Scale Image Recognition.
● Its main contribution was in showing that the depth of the
network is a critical component for good performance.
● Their final best network contains 16 CONV/FC layers and,
appealingly, features an extremely homogeneous
architecture that only performs 3x3 convolutions and 2x2
pooling from the beginning to the end.
-- quoted from CS231N
GoogleNet (2015 by Szegedy et al.) The winner in ILSVRC 2014
Its main contribution was the development of an
Inception Module and the using Average
Pooling instead of Fully Connected layers at the
top of the ConvNet, which dramatically reduced
the number of parameters in the network (4M,
compared to AlexNet with 60M).
-- edited from CS231N
Tip on ConvNets:
Usually, most computation is
spent on convolutions, while
most space is spent on fully
connected layers.
An Inception Module: a new building block..
ResNet (2016 by Kaiming He et al.) The winner in ILSVRC 2015
ResNet (2016 by Kaiming He et al.)
● Deeper network hard to train: Use
skip connections for residual
learning.
● Heavy use of batch normalization.
152 layers..
● No fully connected layers.
Classification:
dog
Segmentation:
Learning Deconvolution Network for
Semantic Segmentation
Up convolution/Convolution transpose/Deconvolution
If you know how to compute gradients in
convolution layers, you know upconv.
Up convolution/Convolution transpose/Deconvolution
x w
11 12 13 14 y
11 12 13
11 12
21 22 23 24
21 22 23
21 22
31 32 33 34
31 32 33
41 42 43 44
Up convolution/Convolution transpose/Deconvolution
x w
11 12 13 14 y
11 12 13
11 12
21 22 23 24
21 22 23
21 22
31 32 33 34
31 32 33
41 42 43 44
Upconvolution/Convolution transpose/Deconvolution
Convolution with stride =>
Upconvolution with input upsampling
See https://github.com/vdumoulin/conv_arithmetic for examples
Dilated Convolution
Fully convolutional network (FCN) variations
Output scores Output scores Output scores
HxWxN HxWxN HxWxN
upsample
upconv
Skip links
dilated
conv
conv
conv
Input image Input image Input image
HxWx3 HxWx3 HxWx3
Dilated/Atrous Convolution
Issues with convolution in dense prediction (image segmentation)
● Use small kernels
○ Receptive field grows linearly with #layers: l∗(k−1)+k
● Use large kernels
○ loss of resolution
Dilated convolutions support exponentially expanding
receptive fields without losing resolution or coverage.
Receptive field: 3 Receptive field: 7 Receptive field: 15
L1: dilation=1 L2: dilation=2 L2: dilation=4
dilation=2 Fig from ICLR 16 paper by Yu and Koltun.
Dilated/Atrous Convolution
Baseline: conv + FC Dilated conv
Fig from ICLR 16 paper by Yu and Koltun.
Outline
● Motivation
● A Simple Neural Network
● Ideas in Deep Net Architectures
● Ideas in Deep Net Optimization
● Practicals and Resources
Optimization
Basics: Gradient descent, SGD, mini-batch SGD, Momentum,
Adam, learning rate decay
Other Ingredients: Data augmentation, Regularization,
Dropout, Xavier initialization, Batch normalization
NN Optimization:
Back Propagation [Hinton et al. 1985]
Gradient Descent with Chain Rule Rebranded.
Fig from Deep Learning by LeCun,
Bengio and Hinton. Nature 2015
SGD, Momentum, RMSProp, Adagrad, Adam
● Batch gradient descent (GD):
○ Update weights once after looking at all the training data.
● Stochastic gradient descent (SGD):
○ Update weights for each sample.
● Mini-batch SGD:
○ Update weights after looking at every “mini batch” of data, say 128 samples.
Let x be the weight/parameters, dx be the gradient of x. In mini-batch, dx is the average within a batch.
SGD (the vanilla update)
where learning_rate is a hyperparameter - a fixed constant. From CS231N
SGD, Momentum, RMSProp, Adagrad, Adam
Momentum:
Initializing the parameters with random numbers
is equivalent to setting a particle with zero initial
velocity at some location.
The optimization process can then be seen as
equivalent to the process of simulating the
parameter vector (i.e. a particle) as rolling on
the landscape.
From CS231N
SGD, Momentum, RMSProp, Adagrad, Adam
Per-parameter adaptive learning rate methods
Adagrad by Duchi et al.:
weights with high gradients =>
effective learning rate reduced
RMSProp by Hinton:
Use moving average to reduce
Adagrad’s aggressive, monotonically
decreasing learning rate
Adam by Kingma et al.:
Use smoothed version of gradients
compared with RMSProp. Default
optimizer (along with Momentum).
From CS231N
Annealing the learning rate (the dark art...)
From Martin Gorner
Annealing the learning rate (the dark art...)
From Martin Gorner
Annealing the learning rate (the dark art...)
● Stairstep decay: Reduce the learning rate by some factor
every few epochs. E.g. half the learning rate every 10
epochs.
● Exponential decay: learning_rate = initial_lr * exp(-kt)
where t is current step.
● “On-demand” decay: Reduce the learning rate when error
plateaus
Optimization
Basics: Gradient descent, SGD, mini-batch SGD, Momentum,
Adam, learning rate decay
Other Ingredients: Data augmentation, Regularization,
Dropout, Xavier initialization, Batch normalization
Dealing with Overfitting: Data Augmentation
Flipping, random crop, random translation, color/brightness change, adding noise...
Pictures from CS231N
Dealing with Overfitting: Regularization, Dropout
L1/L2 regularization on weights: limit the network capacity by encouraging
distributed and sparse weights. When combining L1 and L2 regularization, it’s called
elastic net regularization:
Dropout by Srivastava et al.:
During testing there is no dropout applied,
with the interpretation of evaluating an averaged
prediction across the exponentially-sized
ensemble of all sub-networks.
Applying dropout
during training
Xavier and MSR Initialization
W = 0.01* np.random.randn(D,H)
Problem with random Gaussian initialization: the distribution
of the outputs has a variance that grows with the number of
inputs => Exploding/diminishing output in very deep network.
w = np.random.randn(n) / sqrt(n).
w = np.random.randn(n) * sqrt(2/n).
Data “whitening”
From Martin Gorner
Data “whitening”
From Martin Gorner
Batch Normalization
From Martin Gorner
Batch Normalization
From Martin Gorner
Batch Normalization
From Martin Gorner
Outline
● Motivation
● A Simple Neural Network
● Ideas in Deep Net Architectures
● Ideas in Deep Net Optimization
● Practicals and Resources
Image from Martin Gorner
Data Collecting, Cleaning, Preprocessing > 50%
time
“OS” of Machine/Deep Learning
Caffe, Theano, Torch, Tensorflow, Pytorch, MXNET, …
Matlab in the earlier days. Python and C++ is the popular
choice now.
Deep network debugging, Visualizations
Resources
Stanford CS231N: Convolutional Neural Networks for Visual Recognition
Stanford CS224N: Natural Language Processing with Deep Learning
Berkeley CS294: Deep Reinforcement Learning
Learning Tensorflow and deep learning, without a PhD
Udacity and Coursera classes on Deep Learning
Book by Goodfellow, Bengio and Courville: http://www.deeplearningbook.org/
Talk by LeCun 2013: http://www.cs.nyu.edu/~yann/talks/lecun-ranzato-icml2013.pdf
Talk by Hinton, Bengio, LeCun 2015:
https://www.iro.umontreal.ca/~bengioy/talks/DL-Tutorial-NIPS2015.pdf
What’s not covered...
Sequential Models (RNN, LSTM, GRU)
Deep Reinforcement Learning
3D Deep Learning (MVCNN, 3D CNN, Spectral CNN, NN on Point Sets)
Generative and Unsupervised Models (AE, VAE, GAN etc.)
Theories in Deep Learning
...
Summary
● Why Deep Learning
● A Simple Neural Network
○ Model, Loss and Optimization
● Ideas in deep net architectures
○ Building blocks: FC, ReLU, conv, pooling, unpooling, upconv, dilated conv
○ Classics: MLP, LeNet, AlexNet, NIN, VGG, GoogleNet, ResNet
● Ideas in deep net optimization
○ Basics: GD, SGD, mini-batch SGD, Momentum, Adam, learning rate decay
○ Other Ingredients: Data augmentation, Regularization, Dropout, Batch
normalization
● Practicals and Resources
