Introduction to Deep Learning
TA: Drew Hudson
May 8, 2020
Slides credits: Atharva Parulekar, Jingbo Yang, Drew Hudson, Guanzhi Wang
Overview
● Motivation for deep learning
● Convolutional neural networks
● Recurrent neural networks
● Deep learning tools
But we learned multi-layer perceptron in class?
Expensive to learn. Will not generalize well
Does not exploit the order and local relations in the data!
64x64x3=12288 parameters
We also want many layers
What are areas of deep learning?
Convolutional NN Recurrent NN
Image Time Series
Graph NN
Networks/Relational
Deep RL
Control System
What are areas of deep learning?
Convolutional
Neural Network
Recurrent NN Deep RL Graph NN
Let us look at images in detail
Filters
Why not extract features using filters?
Better, why not let the data dictate
what filters to use?
Learnable filters!!
Convolution on multiple channels
Images are generally RGB !!
How would a filter work on a image
with RGB channels?
The filter should also have 3
channels.
Now the output has a channel for
every filter we have used.
Parameter sharing
Lesser the parameters less computationally intensive the training. This is a
win win as we are reusing parameters.
Translational invariance
Since we are training filters to
detect cats and the moving
these filters over the data, a
differently positioned cat will
also get detected by the same
set of filters.
Filters? Layers of filters?
Images that maximize filter outputs at certain How deeper layers can learn deeper
layers. We observe that the images get more embeddings. How an eye is made up of multiple
complex as filters are situated deeper curves and a face is made up of two eyes.
How do we use convolutions?
Let convolutions extract features and let normal cnn’s decide on them.
Image credit: LeCun et al. (1998)
Convolution really is just a linear operation
In fact convolution is a giant matrix
multiplication.
We can expand the 2 dimensional
image into a vector and the conv
operation into a matrix.
Nonlinearities/Activations
● hidden
For layers, often
● ReLU:
● Hyperbolic tangent:
For output layers, often
● Linear (identity):
● Sigmoid:
● Softmax: (normalize the logits into a discrete probability distribution)
How do we learn?
Instead of
They are “optimizers”
● Momentum: Gradient + Momentum
● Nestrov: Momentum + Gradients
● Adagrad: Normalize with sum of sq
● RMSprop: Normalize with moving
avg of sum of squares
● ADAM: RMsprop + momentum
● https://ruder.io/optimizing-gradient-d
escent/
Mini-batch Gradient Descent
Expensive to compute gradient for large dataset
Memory size
Compute time
Mini-batch: takes a sample of training data
How to we sample intelligently?
Is deeper better?
Deeper networks seem to be
more powerful but harder to train.
● Loss of information during
forward propagation
● Loss of gradient info during
back propagation
There are many ways to “keep
the gradient going”
Solution
Connect the layers, create a gradient highway or information
highway.
ResNet (2015)
Image credit: He et al. (2015)
Initialization
Can we initialize all neurons to zero? Relu units once knocked out and their
output is zero, their gradient flow also
If all the weights are same we will not becomes zero.
be able to break symmetry of the
network and all filters will end up We need small random numbers at
learning the same thing. initialization.
Large numbers, might knock relu units Variance : 1/sqrt(n)
out. Mean: 0
Popular initialization setups
(Xavier, Kaiming) (Uniform, Normal)
Dropout
What does cutting off some network
connections do?
Trains multiple smaller networks in an
ensemble.
Can drop entire layer too!
Acts like a really good regularizer
Tricks for training
Data augmentation if your data set is
smaller. This helps the network
generalize more.
Early stopping if training loss goes
above validation loss.
Random hyperparameter search or grid
search?
CNN sounds like fun!
What are some other areas of deep learning?
Recurrent NN
Time Series
Convolutional NN Deep RL Graph NN
We can also have 1D architectures (remember this)
CNN works on any data where there is a
local pattern
We use 1D convolutions on DNA
sequences, text sequences and music
notes
But what if time series has causal
dependency or any kind of sequential
dependency?
To address sequential dependency?
Use recurrent neural network (RNN)
Latent Output
Unrolling an RNN
Previous output
One time step
RNN Cell
They are really the same cell,
NOT many different cells like kernels of CNN
How does RNN produce result?
Evolving “embedding”
Result after reading
full sentence
I love CS !
There are 2 types of RNN cells
Store in “long term memory” Response to current input Reset gate Update gate
Response to
current input
Long Short Term Memory (LSTM) Gated Recurrent Unit (GRU)
Recurrent AND deep?
Taking last value
Pay “attention” to
everything
Stacking Attention Model
“Recurrent” AND convolutional?
Temporal convolutional network
Temporal dependency achieved through “one-
sided” convolution
More efficient because deep learning
packages are optimized for matrix
multiplication = convolution
No hard dependency
More? Take CS230, CS236, CS231N, CS224N
Convolutional NN Recurrent NN
Image Time Series
Graph NN
Networks/Relational
Deep RL
Control System
Not today, but take CS234 and CS224W
Convolutional NN Recurrent NN
Image Time Series
Graph NN
Networks/Relational
Deep RL
Control System
Tools for deep learning Specialized
Groups
Popular Tools
$50 not enough! Where can I get free stuff?
Google Colab
Azure Notebook
Free (limited-ish) GPU access
Kaggle kernel???
Works nicely with Tensorflow
Amazon SageMaker?
Links to Google Drive
Register a new Google Cloud account To SAVE money
=> Instant $300??
CLOSE your GPU instance
=> AWS free tier (limited compute)
=> Azure education account, $200? ~$1 an hour
Good luck!
Well, have fun too :D
