Perceptron training rule (delta rule)

wi  wi + wi
where wi =  (t – o) xi

Where:
t = c(x) is target value
o is perceptron output
 is small constant (e.g., 0.1) called learning rate
Can prove it will converge
If training data is linearly separable
Gradient descent
Derivation of gradient descent

◆ Gradient descent
- Error (for all training examples.):

- the gradient of E ( partial differentiating ) :

- direction : steepest increase in E.

- Thus, training rule is as follows.

(The negative sign : the direction that decreases E)

Derivation of gradient descent

where xid denotes the single input

components xi for training example d

- The weight update rule for gradient descent


Gradient descent and delta rule

◆Because the error surface

contains only a single global
minimum, this algorithm will
converge to a weight vector with
minimum error, given a
sufficiently small  is used
Hypothesis Space

- Error of different hypotheses

- For a linear unit with two weights, the hypothesis space H is the wo,w1 plane.
- This error surface must be parabolic with a single global minimum (we
desire a hypothesis with minimum error).
Stochastic approximation to gradient descent

- Stochastic gradient descent (i.e. incremental mode) can sometimes

avoid falling into local minima because it uses the various gradient of E
rather than overall gradient of E.
Summary

Perceptron training rule guaranteed to succeed if

training examples are linearly separable
Sufficiently small learning rate η
Linear unit training rule using gradient descent
Converge asymptotically to min. error hypothesis
(Guaranteed to converge to hypothesis with minimum squared
error )
Multilayer networks
Multilayer networks and
and the
the backpropagation
backpropagation algorithm
algorithm

◆ Speech recognition example of multilayer networks learned

by the backpropagation algorithm
◆ Highly nonlinear decision surfaces
Sigmoid Threshold Unit
The Backpropagation algorithm
Adding Momentum

◆ Often include weight momentum α

- nth iteration update depend on (n-1)th iteration

-  : constant between 0 and 1 (momentum)
◆ Roles of momentum term
◆ The effect of keeping the ball rolling through small local
minima in the error surface
◆ The effect of gradually increasing the step size of the
search in regions (greatly improves the speed of learning)
Convergence and Local Minima

◆ Gradient descent to some local minimum

Perhaps not global minimum...
Add momentum
Stochastic gradient descent
Expressive Capabilities of ANNs
 Hidden layer
representations
Hidden layer representations
- This 8x3x8 network was trained to learn the identity function.
- 8 training examples are used.
- After 5000 training iterations, the three hidden unit values encode
the eight distinct inputs using the encoding shown on the right.
Learning the 8x3x8 network
- Most of the interesting weight
changes occurred during the
first 2500 iterations.
 Generalization,
Overfitting, and Stopping
Criterion
Termination condition
Until the error E falls below some predetermined threshold
Techniques to address the overfitting problem
•Weight decay : Decrease each weight by some small factor during each
iteration.
•Cross-validation (k-fold cross-validation)
 Neural Nets for Face Recognition

Training images : 20 different persons with 32 images per person.

After 260 training images, the network achieves an accuracy of 90%
over test set.
Algorithm parameters : η=0.3, α=0.3
Alternative Error Functions

Penalize large weights: (weight decay) : Reducing the risk of

overfitting

Train on target slopes as well as values:

Minimizing the cross entropy : Learning a probabilistic output

function (chapter 6)

−  t d log od + (1 − t d ) log(1 − od )
d∈ D
 Recurrent
Networks
(a) (b) (c)

(a) Feedforward network

(b) Recurrent network
(c) Recurrent network unfolded
in time
 Dynamically
Modifying Network
Structure
To improve generalization accuracy and training efficiency
Cascade-Correlation algorithm (Fahlman and Lebiere 1990)
Start with the simplest possible network (no hidden units) and add complexity
Lecun et al. 1990
Start with the complex network and prune it as we find that certain connectives are
inessential.
 Unsupervised Learning

We can include additional structure in the network so that

the net is forced to make a decision as to which one unit
will respond.

The mechanism by which it is achieved is called

competition.

It can be used in unsupervised learning.

A common use for unsupervised learning is clustering

based neural networks.
 Unsupervised Learning

In a clustering net, there are as many units as the input

vector has components.

Every output unit represents a cluster and the number of

output units limit the number of clusters.

During the training, the network finds the best matching

output unit to the input vector.

The weight vector of the winner is then updated according

to learning algorithm.
 Kohonen Learning

A variety of nets use Kohonen Learning

New weight vector is the linear combination of old
weight vector and the current input vector.
The weight update for cluster unit (output unit) j can be
calculated as:

the learning rate alpha decreases as the learning

process proceeds.
 Kohonen SOM (Self
Organizing Maps)

Since it is unsupervised environment, so the name is Self

Organizing Maps.

Self Organizing NNs are also called Topology Preserving

Maps which leads to the idea of neighborhood of the
clustering unit.

During the self-organizing process, the weight vectors of

winning unit and its neighbors are updated.
 Kohonen SOM (Self
Organizing Maps)

Normally, Euclidean distance measure is used to find the

cluster unit whose weight vector matches most closely to
the input vector.

For a linear array of cluster units, the neighborhood of

radius R around cluster unit J consists of all units j such
that:
 Kohonen SOM (Self
Organizing Maps)

Architecture of SOM
 Kohonen SOM (Self
Organizing Maps)

Structure of Neighborhoods
 Kohonen SOM (Self
Organizing Maps)

Structure of Neighborhoods
 Kohonen SOM (Self
Organizing Maps)

Structure of Neighborhoods
 Kohonen SOM (Self
Organizing Maps)
Neighborhoods do not wrap around from one side of the
grid to other side which means missing units are simply
ignored.
Algorithm:
 Kohonen SOM (Self
Organizing Maps)

Algorithm:

Radius and learning rates may be decreased after each

epoch.
Learning rate decrease may be either linear or
geometric.
 KOHONEN SELF ORGANIZING MAPS

Architecture

neuron i
Kohonen layer
wi

Winning neuron

Input vector X
X=[x1,x2,…xn]  Rn
wi=[wi1,wi2,…,win]  Rn
 Kohonen SOM (Self
Organizing Maps)

Example
Kohonen SOM (Self
Organizing Maps)
Kohonen SOM (Self
Organizing Maps)
Convolutional Neural Network
 History
• In 1995, Yann LeCun
and Yoshua Bengio
introduced the concept
of convolutional neural
networks.
“Deep Learning doesn’t do different things,
it does things differently”
Performance vs Sample Size

Performance

Traditional ML algorithms

Size of Data
 Outline
Supervised Learning
Convolutional Neural Network
Sequence Modelling: RNN and its extensions
Unsupervised Learning
Autoencoder
Stacked Denoising Autoencoder
• Unsupervised Learning (+Supervised)
Generative Adversarial Networks
Reinforcement Learning
Deep Reinforcement Learning
Outline

GAN

Produce Poetry like Shakespeare

Output
Generator
Generated
Input

Network
Shakespeare Fake
Poetry Discriminator
Network
Real Real
Shakespeare
Poetry
Supervised Learning
Traditional pattern recognition models work with hand crafted
features and relatively simple trainable classifiers.

Trainable
Extract Hand Output
Classifier
Crafted (e.g. Outdoor
(e.g. SVM,
Features Yes or No)
Random
Limitations Forrest)

Limitations

Very tedious and costly to develop hand crafted features.

The hand-crafted features are usually highly dependents on one
application.
Deep Learning
Deep learning has an inbuilt automatic multi stage feature learning
process that learns rich hierarchical representations (i.e. features).

Low Level Mid Level High Output

Trainable
Features Features Level (e.g. outdoor,
Classifier indoor)
Features
Deep Learning

Low Level Mid Level High

Trainable Output
Input Features Features Level
Features
Classifier

Image
Pixel Edge Texture Motif Part Object
Text
Character Word Word-group Clause Sentence Story

•Each module in Deep Learning transforms its input

representation into a higher-level one, in a way similar to
human cortex.
Let us see how it all
works!
A Simple Neural Network
An Artificial Neural Network is an information processing paradigm
that is inspired by the biological nervous systems, such as the
human brain’s information processing mechanism.

x1 a1(1)

x2 a2(1)
a1(2) Y
x3 a3(1)

x4 a4(1)

Input Hidden Layers Output

A Simple Neural Network
Softmax 1
w1
x1 a1(1)
1 + 𝑒 −𝑤∗𝑎1(2)
w2
x2 a2(1)
w3 Y
a1(2)

x3 a3(1)
w4

x4 a4(1)

𝑎1(1) = 𝑓 𝑤1 ∗ 𝑥1 + 𝑤2 ∗ 𝑥2 + 𝑤3 ∗ 𝑥3 + 𝑤4 ∗ 𝑥4

f( ) is activation function: Relu or sigmoid

𝑅𝑒𝑙𝑢: max(0, 𝑥)
𝑎1(1) = 𝑚𝑎𝑥 0, 𝑤1 ∗ 𝑥1 + 𝑤2 ∗ 𝑥2 + 𝑤3 ∗ 𝑥3 + 𝑤4 ∗ 𝑥4
Number of Parameters
Softmax
x1 a1(1)

x2 a2(1)
a1(2) Y
x3 a3(1)

x4 a4(1)

Input Hidden Layers Output

4*4 + 4 +1
If the input is an Image?
x1 a1(1)

x2 a2(1)
a1(2) Y
x3 a3(1)

400 X 400 X 3
a480000(1)

x480000

Input Hidden Layers Output

Number of Parameters
480000*480000 + 480000 +1 = approximately 230 Billion !!!
480000*1000 + 1000 +1 = approximately 480 million !!!
Let us see how convolutional layers
help.
Convolutional Layers
0 1 0
Filter 1 -4 1
0 1 0
1 1 1 1 1 10.0156860.0156860.0117650.0156860.0156860.0156860.0156860.9647060.9882350.9647060.8666670.0313730.0235290.007843
0.0078430.741176 1 10.9843140.0235290.0196080.0156860.0156860.0156860.0117650.1019610.972549 1 10.9960780.9960780.9960780.0588240.015686
0.0196080.513726 1 1 10.0196080.0156860.0156860.0156860.0078430.011765 1 1 10.9960780.0313730.0156860.019608 10.011765
0.0156860.733333 1 10.9960780.0196080.0196080.0156860.0156860.0117650.984314 1 10.9882350.0274510.0156860.0078430.007843 10.352941
0.0156860.823529 1 10.9882350.0196080.0196080.0156860.0156860.019608 1 10.9803920.0156860.0156860.0156860.0156860.996078 10.996078
0.0156860.913726 1 10.9960780.0196080.0196080.0196080.019608 1 10.9843140.0156860.0156860.0156860.0156860.952941 1 10.992157
0.0196080.913726 1 10.9882350.0196080.0196080.0196080.0392160.996078 10.0156860.0156860.0156860.0156860.996078 1 1 10.007843
0.0196080.898039 1 10.9882350.0196080.0156860.0196080.9686280.9960780.9803920.0274510.0156860.0196080.9803920.972549 1 1 10.019608
0.0431370.905882 1 1 10.0156860.0352940.968628 1 10.023529 10.7921570.996078 1 10.9803920.9921570.0392160.023529
1 1 1 1 10.9921570.992157 1 10.9843140.0156860.0156860.8588240.996078 10.9921570.5019610.0196080.0196080.023529
0.9960780.992157 1 1 10.9333330.0039220.996078 10.988235 10.992157 1 1 10.988235 1 1 1 1
0.015686 0.74902 1 10.9843140.0196080.0196080.0313730.9843140.0235290.0156860.015686 1 1 1 00.0039220.0274510.980392 1
0.0196080.023529 1 1 10.0196080.0196080.5647060.8941180.0196080.0156860.015686 1 1 10.0156860.0156860.015686 0.05098 1
0.0156860.015686 1 1 10.0470590.0196080.9921570.0078430.0117650.0117650.015686 1 1 10.0156860.0196080.9960780.0235290.996078
0.0196080.0156860.243137 1 10.9764710.035294 10.0039220.0117650.0117650.015686 1 1 10.9882350.988235 10.0039220.015686
0.0196080.0196080.027451 1 10.9921570.2235290.6627450.0117650.0117650.0117650.015686 1 1 10.0156860.0235290.9960780.0117650.011765
0.0156860.0156860.011765 1 1 1 10.0352940.0117650.0117650.0117650.015686 1 1 10.0156860.0156860.9647060.0039220.996078
0.0078430.0196080.0117650.054902 1 10.9882350.0078430.0117650.0117650.0156860.011765 1 1 10.0156860.0156860.0156860.023529 1
0.0078430.0078430.0156860.0156860.960784 10.4901960.0156860.0156860.0156860.0078430.027451 1 1 10.0117650.0117650.043137 1 1
0.0235290.0039220.0078430.0235290.9803920.9764710.0392160.0196080.0078430.0196080.015686 1 1 1 1 1 1 1 1 1

Input Image Convoluted Image

▪ Inspired by the neurophysiological experiments conducted by Hubel and Wiesel 1962.

Convolutional Layers
What is Convolution?
ℎ1 = 𝑓 𝑎 ∗ 𝑤1 + 𝑏 ∗ 𝑤2 + 𝑒 ∗ 𝑤3 + 𝑓 ∗ 𝑤4

a b c d w1 w2
h1 h2
e f g h w3 w4
i j k l
m n o p

Input Image Filter Convolved Image

(Feature Map)

ℎ2 = 𝑓 𝑏 ∗ 𝑤1 + 𝑐 ∗ 𝑤2 + 𝑓 ∗ 𝑤3 + 𝑔 ∗ 𝑤4

Number of Parameters for one feature map = 4

Number of Parameters for 100 feature map = 4*100
Lower Level to More Complex Features

w1 w2

w3 w4
w5 w6

w7 w8
Filter 1
Filter 2

Input Image
Layer 1 Layer 2
Feature Map Feature Map
▪ In Convolutional neural networks, hidden units are only connected to local receptive field.
Pooling
Max pooling: reports the maximum output within a rectangular
neighborhood.
Average pooling: reports the average output of a rectangular
neighborhood.

MaxPool with 2X2 filter with

1 3 5 3 stride of 2
4 2 3 1
4 5
3 1 1 3
3 4
0 1 0 4

Input Matrix Output Matrix

Convolutional Neural Network
Maxpool
Output
Feature Extraction Architecture Vector

Living Room

Bed Room
128

256

512

512
256

512

512
128

256

512
512
Kitchen
64
64

Bathroom

Outdoor
Max Pool
Filter

Fully Connected
Layers
Convolutional Neural Networks
Output: Binary, Multinomial, Continuous, Count
Input: fixed size, can use padding to make all images same
size.
Architecture: Choice is ad hoc
requires experimentation.
Optimization: Backward propagation
hyper parameters for very deep model can be estimated properly only if you
have billions of images.
Use an architecture and trained hyper parameters from other papers
(Imagenet or Microsoft/Google APIs etc)
Computing Power: Buy a GPU!!
Automatic Colorization of Black and White Images
Optimizing Images

Post Processing Feature Optimization

(Color Curves and Details)

Post Processing Feature Optimization Post Processing Feature Optimization

(Illumination) (Color Tone: Warmness)
 Convolution
1D (continuous, discrete) : Input

 Kernel
f * g (x) =  f ( )g(x − )d 
 =−
N −1 Output is
=  f ( )g(x − ) sometimes called
 =0 Feature map
2D (continuous, discrete) :
 
f * g (x, y) =   f ( ,  )g(x − , y −  )d d 
 =−  =−
N −1 N −1
=   f ( ,  )g(x − , y −  )
 =0  =0
 Convolution Properties
• Commutative:
f*g = g*f
• Associative:
(f*g)*h = f*(g*h)
• Homogeneous:
f*(g)=  f*g
• Additive (Distributive):
f*(g+h)= f*g+f*h
• Shift-Invariant
f*g(x-x0,y-yo)= (f*g) (x-x0,y-yo)
 ConvNet
• ConvNet architectures for images:
– fully-connected structure does not scale to large
images
– the explicit assumption that the inputs are images
– allows us to encode certain properties into the
architecture.
– These then make the forward function more efficient
to implement
– Vastly reduce the amount of parameters in the
network.
• 3D volumes: neurons arranged in 3 dimensions:
width, height, depth.
 Convnets

Layers used to build ConvNets:

• a stacked sequence of
layers. 3 main types
• Convolutional Layer,
Pooling Layer, and Fully-
Connected Layer • every layer of a ConvNet
transforms one volume of
activations to another through a
differentiable function.
 The replicated feature approach
• Use many different copies of the
same feature detector with The red connections all
differentpositions. have the same weight.
– Could also replicate across scale and
orientation (tricky and expensive)
– Replication greatly reduces the
number of free parameters to be
learned.
• Use several different feature types,
each with its own map of
replicated detectors.
– Allows each patch of image to be
represented in several ways.
 Backpropagation with weight constraints

• It’s easy to modify the To constrain : w1 = w2

backpropagation algorithm to we need : w1 = w2
incorporate linear constraints
between the weights. E E
compute: and
• We compute the gradients as w1 w2
usual, and then modify the
gradients so that they satisfy E + E
use for w1and w2
the constraints. w1 w2
– So if the weights started off
satisfying the constraints, they
will continue to satisfy them.
 What does replicating the feature detectors achieve?
• Equivariant activities: Replicated features do not make the
neural activities invariant to translation. The activities are
equivariant.
representation translated
by active representation
neurons

translated
image image

• Invariant knowledge: If a feature is useful in some locations

during training, detectors for that feature will be available in
all locations during testing.
 Pooling the outputs of replicated feature
detectors
• Get a small amount of translational invariance at each level
by averaging four neighboring replicated detectors to give a
single output to the next level.
– This reduces the number of inputs to the next layer of feature
extraction, thus allowing us to have many more different feature
maps.
– Taking the maximum of the four works slightly better.
• Problem: After several levels of pooling, we have lost
information about the precise positions of things.
– This makes it impossible to use the precise spatial relationships
between high-level parts for recognition.
 Example Architecture for CIFAR-10
• [INPUT - CONV - RELU - POOL - FC]
• INPUT [32x32x3] : the raw pixel values of the image
• CONV will compute the output of neurons that are connected to
local regions in the input. With 12 filters, the output volume is
[32x32x12]
• RELU : apply an elementwise activation function, such as the
max(0,x)
• POOL will perform a downsampling operation along the spatial
dimensions (width, height), resulting in volume such as [16x16x12].
• FC layer will compute the class scores, resulting in volume of size
[1x1x10], where each of the 10 numbers correspond to a class
score, such as among the 10 categories of CIFAR-10
 Convolution Layer
• The Conv layer is the core building block of a CNN
• The parameters consist of a set of learnable filters.
• Every filter is small spatially (width and height), but extends through the
full depth of the input volume, eg, 5x5x3
• During the forward pass, we slide (convolve) each filter across the width
and height of the input volume and compute dot products between the
entries of the filter and the input at any position.
• produce a 2-dimensional activation map that gives the responses of that
filter at every spatial position.
• Intuitively, the network will learn filters that activate when they see some
type of visual feature
• A set of filters in each CONV layer
– each of them will produce a separate 2-dimensional activation map
– We will stack these activation maps along the depth dimension and produce
the output volume.
Convolutional Neural Network 2
Convolution
 Convolutions: More detail

32x32x3 image

32 height

32 width
3 depth
 Convolutions: More detail
32x32x3 image

5x5x3 filter
32

Convolve the filter with the image

i.e. “slide over the image spatially,
computing dot products”

32
3
Convolutions: More detail
Convolution Layer
32x32x3 image
5x5x3 filter
32

1 number:
the result of taking a dot product between the
filter and a small 5x5x3 chunk of the image
32 (i.e. 5*5*3 = 75-dimensional dot product + bias)
3
Convolutions: More detail
Convolution Layer
activation map
32x32x3 image
5x5x3 filter
32

28

convolve (slide) over all

spatial locations

32 28
3 1
Convolutions: More detail
consider a second, green filter
Convolution Layer
32x32x3 image activation maps
5x5x3 filter
32

28

convolve (slide) over all

spatial locations

32 28
3 1
Convolutions: More detail
For example, if we had 6 5x5 filters, we’ll get 6 separate activation maps:
activation maps

32

28

Convolution Layer

32 28
3 6

We stack these up to get a “new image” of size 28x28x6!

Convolutions: More detail
Preview: ConvNet is a sequence of Convolution Layers, interspersedwith
activation functions

32 28

CONV,
ReLU
e.g. 6
5x5x3
32 filters 28
3 6
Convolutions: More detail
Preview: ConvNet is a sequence of Convolutional Layers, interspersed with activation
functions

32 28 24

….
CONV, CONV, CONV,
ReLU ReLU ReLU
e.g. 6 e.g. 10
5x5x3 5x5x6
32 filters 28 filters 24
3 6 10
Convolutions: More detail
[From recentYann
Preview LeCun slides]
Convolutions: More detail
one filter =>
one activation map example 5x5 filters
(32 total)

We call the layer convolutional

because it is related to convolution
of two signals:

elementwise multiplication and sum of

a filter and the signal(image)
Convolutions: More detail
A closer look at spatial dimensions:
activation map
32x32x3 image
5x5x3 filter
32

28

convolve (slide) over all

spatial locations

32 28
3 1
Convolutions: More detail
A closer look at spatial dimensions:

• 7
• 7x7 input
(spatially)
assume 3x3
filter

• 7
Convolutions: More detail
A closer look at spatial dimensions:

• 7
• 7x7 input
(spatially)
assume 3x3
filter

• 7
Convolutions: More detail
A closer look at spatial dimensions:

• 7
• 7x7 input
(spatially)
assume 3x3
filter

• 7
Convolutions: More detail
A closer look at spatial dimensions:

• 7
• 7x7 input
(spatially)
assume 3x3
filter

• 7
Convolutions: More detail
A closer look at spatial dimensions:

• 7
• 7x7 input (spatially)
assume 3x3 filter
7 => 5x5 output
Convolutions: More detail
A closer look at spatial dimensions:

7
7x7 input (spatially)
assume 3x3 filter
applied with stride 2

7
Convolutions: More detail
A closer look at spatial dimensions:

7
7x7 input (spatially)
assume 3x3 filter
applied with stride 2

7
Convolutions: More detail
A closer look at spatial dimensions:

7
7x7 input (spatially)
assume 3x3 filter
applied with stride 2
=> 3x3 output!
7
Convolutions: More detail
A closer look at spatial dimensions:

7
7x7 input (spatially)
assume 3x3 filter
applied with stride 3?

7
Convolutions: More detail
A closer look at spatial dimensions:

7
7x7 input (spatially)
assume 3x3 filter
applied with stride 3?

7 doesn’t fit!
cannot apply 3x3 filter on
7x7 input with stride 3.
Convolutions: More detail
N
Output size:
(N - F) / stride + 1
F
e.g. N = 7, F = 3:
F N
stride 1 => (7 - 3)/1 + 1 = 5
stride 2 => (7 - 3)/2 + 1 = 3
stride 3 => (7 - 3)/3 + 1 = 2.33 :\
Convolutions: More detail
In practice: Common to zero pad the border
0 0 0 0 0 0
e.g. input 7x7
0
3x3 filter, applied with stride 1
0 pad with 1 pixel border => what is the output?
0

(recall:)
(N - F) / stride + 1
Convolutions: More detail
In practice: Common to zero pad the border
0 0 0 0 0 0
e.g. input 7x7
0
3x3 filter, applied with stride 1
0 pad with 1 pixel border => what is the output?
0

0
7x7 output!
Convolutions: More detail
In practice: Common to zero pad the border
0 0 0 0 0 0
e.g. input 7x7
0
3x3 filter, applied with stride 1
0 pad with 1 pixel border => what is the output?
0

0
7x7 output!
in general, common to see CONV layers with
stride 1, filters of size FxF, and zero-padding with
(F-1)/2. (will preserve size spatially)
e.g. F = 3 => zero pad with 1
F = 5 => zero pad with 2
F = 7 => zero pad with 3

(N + 2*padding - F) / stride + 1
Convolutions: More detail
Examples time:

Input volume: 32x32x3

10 5x5 filters with stride 1, pad 2

Output volume size: ?

Convolutions: More detail
Examples time:

Input volume: 32x32x3

10 5x5 filters with stride 1, pad 2

Output volume size:

(32+2*2-5)/1+1 = 32 spatially, so
32x32x10
Convolutions: More detail
Examples time:

Input volume: 32x32x3

10 5x5 filters with stride 1, pad 2

Number of parameters in this layer?

Convolutions: More detail
Examples time:

Input volume: 32x32x3

10 5x5 filters with stride 1, pad 2

Number of parameters in this layer?

each filter has 5*5*3 + 1 = 76 params (+1 for bias)
=> 76*10 = 760
Convolutions: More detail
 Spatial arrangement
• Three hyperparameters control the size of the
output volume
– Depth: no of filters, each learning to look for
something different in the input.
– the stride with which we slide the filter.
– pad the input volume with zeros around the
border.
 Spatial arrangement
• We compute the spatial size of the output
volume as a function of
– the input volume size (W)
– the receptive field size of the Conv Layer neurons (F)
– the stride with which they are applied (S)
– the amount of zero padding used (P) on the border.
• The number of neurons that “fit” is given by
(W−F+2P)/(S+1)
– For a 7x7 input and a 3x3 filter with stride 1 and pad 0
we would get a 5x5 output.
– With stride 2 we would get a 3x3 output.
• one spatial dimension (x-axis), one neuron with a receptive field
size of F = 3, the input size is W = 5, and zero padding of P = 1
• Stride = 1, 2

• The Krizhevsky et al. architecture that won the ImageNet 2012

• images of size [227x227x3].
• the first Convolutional Layer, used neurons with receptive field size F=11,
stride S=4, no zero padding P=0
• Since (227 - 11)/4 + 1 = 55, the Conv layer had a depth of K=96,
• the Conv layer output volume had size [55x55x96].
• Each of the 55*55*96 neurons in this volume was connected to a region of
size [11x11x3] in the input volume.
• Moreover, all 96 neurons in each depth column are connected to the same
[11x11x3] region of the input,
 Parameter Sharing
• Parameter sharing controls the number of parameters.
• If there are 55*55*96 = 290,400 neurons in the first Conv Layer, and
each has 11*11*3 = 363 weights and 1 bias. Together, this adds up
to 290400 * 364 = 105,705,600 parameters on the first layer of the
ConvNet alone.
• Reduce by parameter sharing
• now have only 96 unique set of weights (one for each depth slice),
for a total of 96*11*11*3 = 34,848 unique weights, or 34,944
parameters (+96 biases)
• During backpropagation, every neuron in the volume will compute
the gradient for its weights, but these gradients will be added up
across each depth slice and only update a single set of weights per
slice.
• Example filters learned by Krizhevsky.
• 96 filters each of size [11x11x3], each is
shared by the 55*55 neurons in one depth
slice.
 Summary of Conv Layer
• Accepts a volume of size W1×H1×D1
• Requires four hyperparameters:
– Number of filters K
– their spatial extent F
– the stride S
– the amount of zero padding P
• Produces a volume of size W2×H2×D2
– W2=(W1−F+2P)/S+1
– H2=(H1−F+2P)/S+1
– D2=K
• With parameter sharing, it introduces F⋅F⋅D1 weights per filter, for a total
of (F⋅F⋅D1)⋅K weights and K biases.
• In the output volume, the d-th depth slice (of size W2×H2) is the result of
performing a valid convolution of the d-th filter over the input volume
with a stride of S, and then offset by d-th bias.
 Spatial Pooling
• Sum or max over non-overlapping / overlapping regions
• Role of pooling:
• Invariance to small transformations
• Larger receptive fields (neurons see more of input)

Max

Sum
 3. Spatial Pooling
• Sum or max over non-overlapping / overlapping regions
• Role of pooling:
• Invariance to small transformations
• Larger receptive fields (neurons see more of input)
 Pooling Layer
• Insertion of pooling layer:
– reduce the spatial size of the representation
reduce the amount of parameters and computation in the network, and
hence also control overfitting.
• The Pooling Layer operates independently on every depth slice of
the input and resizes it spatially, using the MAX operation.
• The most common form is a pooling layer with filters of size 2x2
applied with a stride of 2 -- downsamples every depth slice in the
input by 2 along both width and height,
• MAX operation would in take a max over 4 numbers (little 2x2
region in some depth slice).
• The depth dimension remains unchanged.
 General pooling layer
• Accepts a volume of size W1×H1×D1
• Requires two hyperparameters:
– their spatial extent F
– the stride S
• Produces a volume of size W2×H2×D2 where:
– W2=(W1−F)/S+1
– H2=(H1−F)/S+1
– D2=D1
• Introduces zero parameters
• Other pooling functions: Average pooling, L2-
norm pooling
 General pooling

• Backpropagation. the backward pass for a max(x, y) operation

routes the gradient to the input that had the highest value in
the forward pass.
• Hence, during the forward pass of a pooling layer you may
keep track of the index of the max activation (sometimes also
called the switches) so that gradient routing is efficient during
backpropagation.
 Getting rid of pooling
1. Striving for Simplicity: The All Convolutional Net proposes to
discard the pooling layer and have an architecture that only
consists of repeated CONV layers.
• To reduce the size of the representation they suggest using larger
stride in CONV layer once in a while.
• Argument:
– The purpose of pooling layers is to perform dimensionality reduction to
widen subsequent convolutional layers' receptive fields.
– The same effect can be achieved by using a convolutional layer: using a
stride of 2 also reduces the dimensionality of the output and widens the
receptive field of higher layers.
• The resulting operation differs from a max-pooling layer in that
– it cannot perform a true max operation
– it allows pooling across input channels.

Springenberg, Jost Tobias, et al. "Striving for simplicity: The all

convolutional net." arXiv preprint arXiv:1412.6806 (2014).
 Getting Rod of Pooling 2
2. Very Deep Convolutional Networks for Large-Scale Image
Recognition.
• The core idea here is that hand-tuning layer kernel sizes to achieve
optimal receptive fields (say, 5×5 or 7×7) can be replaced by simply
stacking homogenous 3×3 layers.
• The same effect of widening the receptive field is then achieved by
layer composition rather than increasing the kernel size
– three stacked 3×3 have a 7×7 receptive field.
– At the same time, the number of parameters is reduced:
– a 7×7 layer has 81% more parameters than three stacked 3×3 layers.

Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale
image recognition." arXiv preprint arXiv:1409.1556 (2014).
 Fully-connected layer
• Neurons in a fully connected layer have full connections to all
activations in the previous layer
• Their activations can hence be computed with a matrix
multiplication followed by a bias offset.
• Converting FC layers to CONV layers
• the only difference between FC and CONV layers is that the
neurons in the CONV layer are connected only to a local
region in the input, and that many of the neurons in a CONV
volume share parameters.
• However, the neurons in both layers still compute dot
products, so their functional form is identical.
 Converting FC layers to CONV layers
• For any CONV layer there is an FC layer that implements the same forward
function.
• The weight matrix would be a large matrix that is mostly zero except for at
certain blocks (due to local connectivity) where the weights in many of the
blocks are equal (due to parameter sharing).
• Conversely, any FC layer can be converted to a CONV layer.
• For example, an FC layer with K=4096 that is looking at some input volume
of size 7×7×512
• can be equivalently expressed as a CONV layer with F=7,P=0,S=1,K=4096.
• In other words, we are setting the filter size to be exactly the size of the
input volume, and hence the output will simply be 1×1×4096 since only a
single depth column “fits” across the input volume, giving identical result
as the initial FC layer.
 ConvNet Architectures
Layer Patterns
• The most common architecture
• stacks a few CONV-RELU layers,
• follows them with POOL layers,
• and repeats this pattern until the image has been merged spatially
to a small size.
• At some point, it is common to transition to fully-connected layers.
The last fully-connected layer holds the output, such as the class
scores. In other words, the most common ConvNet architecture
follows the pattern:
INPUT -> [[CONV -> RELU]*N -> POOL?]*M ->[FC -> RELU]*K -> FC
• N >= 0 (and usually N <= 3), M >= 0, K >= 0
Prefer a stack of small filter CONV to one large receptive field CONV layer.
three layers of 3x3 CONV vs a single CONV layer with 7x7
receptive fields.
• The receptive field size is identical in spatial extent (7x7), but
with several disadvantages.
1. The neurons would be computing a linear function over the input,
while the three stacks of CONV layers contain non-linearities that
make their features more expressive.
2. If we suppose that all the volumes have C channels, the single 7x7
CONV layer would contain C×(7×7×C)=49C2 parameters, while the
three 3x3 CONV layers would contain 3×(C×(3×3×C))=27C2
parameters.
• Intuitively, stacking CONV layers with tiny filters as opposed to
having one CONV layer with big filters allows us to express
more powerful features of the input, and with fewer
parameters.
 Recent Departures
• The conventional paradigm of a linear list of layers
has recently been challenged, in
1. Google’s Inception architectures
2. current (state of the art) Residual Networks from
Microsoft Research Asia.
• Both of these feature more intricate and different
connectivity structures.
DIABETIC RETINOPATHY
LEARNING OBJECTIVES
• Recognize the importance of diabetic retinopathy as a public
health problem
• Discuss diabetic retinopathy as a leading cause of blindness in
developed countries
• Identify the risk factors for diabetic retinopathy
• Describe and distinguish between the stages of diabetic
retinopathy
• Understand the role of risk factor control and annual dilated eye
exams in the prevention of vision loss
DIABETES MELLITUS
Diabetes Mellitus is a group of diseases characterized by high blood glucose
levels. Diabetes results from defects in the body's ability to produce and/or use
insulin.
• Type 1 diabetes is usually diagnosed in children and young adults, and was
previously known as juvenile diabetes. In type 1 diabetes, the body does not
produce insulin. 5% of people with diabetes have this form of the disease.
• In Type 2 diabetes, either the body does not produce enough insulin or the
cells ignore the insulin. This is the most common form of diabetes.
DIABETIC RETINOPATHY (DR)
DEFINITION
• Progressive dysfunction of the retinal blood vessels
caused by chronic hyperglycemia.
• DR can be a complication of diabetes type 1 or
diabetes type 2.
• Initially, DR is asymptomatic, if not treated though it
can cause low vision and blindness.

http://www.mdconsult .com/da s/boo k/pdf /282715756 -3/978 -0-3 23-04332 -8/4 -u1.0 - B978- 0-323 -04332 -8..00 092-5 ..DOC PD F.pdf? isbn=97 8-0 -323- 04332-8 &e id= 4-u1. 0- B978 -0-323 -04332 -8..0 0092-5 ..DO CPD F
WHAT IS THE RETINA?
• The retina is a multilayered, light sensitive neural tissue
lining the inner eye ball. Light is focused onto the retina
and then transmitted to the brain through the optic
nerve.
• The macula is a highly sensitive area in the center of
the retina, responsible for central vision. The macula is
needed for reading, recognizing faces and executing
other activities that require fine, sharp vision.
RETINA
Healthy Retina Diabetic Retinopathy
DIABETIC RETINOPATHY
EPIDEMIOLOGY

• The total number of people with diabetes

is projected to rise from 285 million in
2010 to 439 million in 2030.
• Diabetic retinopathy is responsible for
1.8 million of the 37 million cases of
blindness throughout the world .
• Diabetic retinopathy (DR) is the leading
cause of blindness in people of working
age in industrialized countries.
Causes of global blindness in millions of people
(WHO 2002)
20
18
16
14
12
10
8
6
4
2
0
DIABETIC RETINOPATHY
EPIDEMIOLOGY

• The best predictor of diabetic retinopathy is the duration

of the disease
• After 20 years of diabetes, nearly 99% of patients with
type 1 diabetes and 60% with type 2 have some degree
on diabetic retinopathy
• 33% of patients with diabetes have signs of diabetic
retinopathy
• People with diabetes are 25 times more likely to become
blind than the general population.
PREVALENCE OF DIABETIC RETINOPATHY AFTER
20 YEARS OF DIAGNOSIS
http://www.who.int/bulletin/volumes/82/11/en/844.pdf
DIABETIC RETINOPATHY SYMPTOMS
Diabetic retinopathy is asymptomatic in early stages of the disease
As the disease progresses symptoms may include
• Blurred vision
• Floaters
• Fluctuating vision
• Distorted vision
• Dark areas in the vision
• Poor night vision
• Impaired color vision
• Partial or total loss of vision
Risk factors

• Duration of diabetes
• Poor Blood Sugar control
• HTN
• Hyperlipidemia
• Barriers to care

http://jama.ama-assn.org/content/304/6/649.short?rss=1
The Effect of Intensive Diabetes Treatment
On the Progression of Diabetic Retinopathy
In Insulin-Dependent Diabetes Mellitus

The Diabetes Control and Complications Trial

The Diabetes Control and Complications Trial Research Group

Intensive control reduced the risk of developing retinopathy by 76%

and slowed progression of retinopathy by 54%; intensive control
also reduced the risk of clinical neuropathy by 60% and albuminuria
by 54%.
RISK FACTORS DIABETIC RETINOPATHY

Duration of diabetes is a major risk

factor associated with the development
of diabetic retinopathy

The severity of hyperglycemia is the

key alterable risk factor associated with
the development of diabetic retinopathy

http://one.aao.org/CE/PracticeGuidelines/PPP_Content.aspx?cid=d0c853d3-219f-487b-a524-326ab3cecd9a
HOW DIABETES CAUSES VISION LOSS
How diabetes cause vision loss

Macular Clinical
significant
edema
macular edema

Preclinical Background Vision

Diabetes changes DR loss

Vitreous hemorrhage
Preproliferative Proliferative and/or Retinal
DR DR detachment and/or
neovascular glaucoma
 PATHOPHYSIOLOGY
Diabetic Retinopathy is a microvasculopathy that
causes:
• Retinal capillary occlusion
• Retinal capillary leakage
MICROVASCULAR OCCLUSION
Microvascular occlusion is caused by:
• Thickening of capillary basement membranes
• Abnormal proliferation of capillary endothelium
• Increased platelet adhesion
• Increased blood viscosity
• Defective fibrinolysis

Retina in systemic disease : a color manual of ophthalmoscopy / Homayoun

Tabandeh, Morton F. Goldberg 2009
 Microvascular
Occlusion

Ischemia

Infarction

Increased VEFG

Cotton – wool spot

Neovascularization

Vitreous Neovascular
Fibrovascular bands
hemorrhage glaucoma

Tractional retinal
detachment Retina in systemic disease : a color manual of
ophthalmoscopy / Homayoun Tabandeh, Morton F.
Goldberg 2009
MICROVASCULAR LEAKAGE
Microvascular leakage is caused by:
• Impairment of endothelial tight junctions
• Loss of pericytes
• Weakening of capillary walls
• Elevated levels of vascular endothelial growth factor (VEGF)

Retina in systemic disease : a color manual of ophthalmoscopy / Homayoun Tabandeh,

Morton F. Goldberg 2009
 Microvascular Leakage

Retinal
Edema Hard exudates
hemorrhage

.
RECOMMENDED
DiabeticEYE EXAMINATION
Eye Disease
SCHEDULE Key Points
Diabetes Type Recommended Time of Recommended Follow-
First Examination up*

Type 1 3-5 years after Yearly

diagnosis

Type 2 At time of diagnosis Yearly

Prior to pregnancy • Treatments

Prior exist but workNobest
to conception retinopathy to mild
(type 1 or type 2) and early in the first moderate NPDR every
before vision
trimesteris lost 3-12 months
Severe NPDR or worse
every 1-3 months.

*Abnormal findings may dictate more frequent follow-up examinations

h ttp://one.aao.org/CE/PracticeGuidelines/PPP_Content.aspx?cid=d0c853d3 -219f-487b-a524-326ab3cecd9a
International Clinical Diabetic Retinopathy Disease Severity
Scale
Findings Observable upon Dilated
Proposed Disease Severity Level Ophthalmoscopy
Findings Obsd
No apparent retinopathy No abnormalities

Mild nonproliferative diabetic retinopathy Microaneurysms only

More than just microaneurysms but less than severe NPDR

Moderate nonproliferative diabetic retinopathy

Any of the following:

Severe nonproliferative diabetic retinopathy More than 20 intraretinal hemorrhages in each of four
quadrants
Definite venous beading in two or more quadrants
Prominent IRMA in one or more quadrants
and no signs of proliferative retinopathy.

One or both of the following:

Proliferative diabetic retinopathy Neovascularization
Vitreous/preretinal hemorrhage
No retinopathy
 MILD NONPROLIFERATIVE
DIABETIC RETINOPATHY
Characteristics
• Microaneurysms only
MILD NONPROLIFERATIVE DIABETIC
RETINOPATHY

Microaneurysms
MODERATE NONPROLIFERATIVE DIABETIC
RETINOPATHY (NPDR)

Characteristics
• More than just microaneurysms but less than severe NPDR but
less than severe NPD
MODERATE NONPROLIFERATIVE DIABETIC
RETINOPATHY (NPDR)

Microaneurysm

Hard exudates

Flamed shaped
hemorrhage
MODERATE NONPROLIFERATIVE
DIABETIC RETINOPATHY (NPDR)

Hard exudates

microaneurysm
SEVERE NONPROLIFERATIVE
DIABETIC RETINOPATHY (NPDR)
Any of the following:
• More than 20 intraretinal hemorrhages in each of four
quadrants
• Definite venous beading in two or more quadrants
• Prominent Intraretinal Microvascular Abnormalities
(IRMA) in one or more quadrants
• And no signs of proliferative retinopathy
Severe Nonproliferative Diabetic Retinopathy
(NPDR)

Venous beading
Proliferative Diabetic Retinopathy (PDR)

Characteristics
• Neovascularization
• Vitreous/preretinal
hemorrhage
PROLIFERATIVE
DIABETIC
RETINOPATHY Cotton-wool
spot

Neovascularization

Neovascularization
Hard exudate
Blot hemorrhage
HIGH-RISK PROLIFERATIVE DIABETIC
RETINOPATHY

At risk for serious vision loss

Any combination of three of the following four findings
• Presence of vitreous or preretinal hemorrhage.
• Presence of new vessels (neovascularization, NV)
• Location of NV on or near the optic disc.
• Moderate to severe extent of new vessels.

Basic and Clinical Science Course, Section 12: Retina and Vitreous AAO
DIABETIC MACULAR EDEMA
• Diabetic macular edema is the leading cause of legal
blindness in diabetics.
• Diabetic macular edema can be present at any stage of
the disease, but is more common in patients with
proliferative diabetic retinopathy.
Meta analysis and review on the effect on bevacizumab id diabetic macular edema
Graefes Arch Clin Exp Ophthalmol(2011) 249:15-27
Why is Diabetic macular edema so important?
• The macula is responsible for central vision.
• Diabetic macular edema may be asymptomatic at
first. As the edema moves in to the fovea (the center
of the macula) the patient will notice blurry central
vision. The ability to read and recognize faces will be
compromised.

Macula
Fovea
Normal Macular Edema
CLINICALLY SIGNIFICANT MACULAR EDEMA
(CSME)
• Thickening of the retina at or within 500 µm of the
center of the macula.
• Hard exudates at or within 500 µm of the center of the
macula, if associated with thickening of the adjacent
retina.
• Area of retinal thickening 1 disc area or larger, within 1
disc diameter of the center of the macula.

ETDRS
INTERNATIONAL CLINICAL DIABETIC MACULAR EDEMA
DISEASE SEVERITY SCALE

Proposed disease severity level Findings observable upon dilated

ophthalmoscopy
DME apparently absent No apparent retinal thickening or hard exudates in
posterior pole

DME apparently present Some apparent retinal thickening or hard exudates in

posterior pole

DME present Mild DME (some retinal thickening or hard exudates in

posterior pole but distant from the center of the
macula)

Moderate DME (retinal thickening or hard

exudates approaching the center of the macula but not
involving the center)

Severe DME (retinal thickening or hard exudates

Proposed International Clinical Diabetic involving the center of the macula)
Retinopathy and Diabetic Macular Edema
Disease Severity Scales
Ophthalmology Volume 110, Number 9, September 2003
Imaging of macular edema with optical
coherence tomography
PREVENTION
90 percent of diabetic eye disease can
be prevented simply by proper regular
examinations, treatment and by
controlling blood sugar.
Primary prevention
Strict glycemic control
Blood pressure control

Secondary prevention
Annual eye exams

Tertiary prevention
Retinal Laser photocoagulation
Vitrectomy
DIABETIC RETINOPATHY TREATMENT

The best measure for prevention of

loss of vision from diabetic
retinopathy is strict glycemic control
LASER PHOTOCOAGULATION
Laser Photocoagulation is recommended for eyes with:
• Clinical significant macular edema CSME
• High risk Proliferative diabetic retinopathy
DIABETIC RETINOPATHY TREATMENT
ONCE DR THREATENS VISION TREATMENTS CAN INCLUDE:

Laser therapy to seal leaking blood vessels

(focal laser)

Laser therapy to reduce retinal oxygen

demand (scatter laser)

Surgical removal of blood from the eye

(vitrectomy)
DIABETIC RETINOPATHY TREATMENT
NEWER DEVELOPMENTS:

The use of anti-vascular endothelial growth

factor antibodies has been shown to be
useful in the treatment of DR

Anti-VEGF antibody treatment appears to

be useful for both macular edema and
proliferative retinopathy

Studies to determine the exact role of anti-

VEGF treatment in relation to laser
treatment in specific situations are
underway.
 CONCLUSIONS

Diabetic Retinopathy is
preventable through strict
glycemic control and annual
dilated eye exams by an
ophthalmologist.
The Guerrilla Eye Service of the UPMC Eye Center is dedicated
to eliminating barriers to eye care for patients in the Western
Pennsylvania area.
Self-driving cars
Question
How would you define a self-driving car?
Definition: What is an autonomous car?
● Autonomous Car: A driverless vehicle capable of fulfilling the main
transportation capabilities of a traditional car.
Classifications of Autonomy according to the NHTSA.
● Level 0: The driver completely controls the vehicle at all times.
● Level 1: Individual vehicle controls are automated, such as electronic stability
control or automatic braking.
● Level 2: At least two controls can be automated in unison, such as adaptive
cruise control in combination with lane keeping.
● Level 3: The driver can fully cede control of all safety-critical functions in certain
conditions and the car provides a "sufficiently comfortable transition time" for the
driver to do so.
● Level 4: The vehicle performs all safety-critical functions for the entire trip, with
the driver not expected to control the vehicle at any time.
Purpose
What kinds of things does a self-driving car need to be able to do?
Purpose
● navigate to a given destination based on passenger-provided instructions

● avoid environmental obstacles

● safely avoid other vehicles

● obey the laws of the road

History
Linrrican Wonder
● Houdina Radio Control, 1925
● Made by Francis P Houdina
● Traveled up Broadway and down Fifth Avenue through the thick of the traffic
jam
Futurama
● sponsored by General Motors at the 1939
World's Fair
● radio-controlled electric cars
○ propelled via electromagnetic fields
RCA Labs
● 1953- RCA Labs built a miniature car guided and controlled by wires
● 1958- Full sized system made
○ developed in collab. with General Motors
Mercedes Benz
● 1980’s- vision-guided Mercedes-Benz robotic van
○ designed by Ernst Dickmanns and his team at the Bundeswehr University
Munich
● achieved a speed of 39 miles per hour (63 km/h) on streets without traffic
History
● Carnegie Mellon’s Navlab and ALV projects in 1984
● Mercedes-Benz and Budeswehr University Munich’s EUREKA Promethius
Project in 1987
● Others:
○ Continental Automotive Systems, IAV, Autoliv Inc., Bosch, Nissan,
Renault, Toyota, Audi, Volvo, Peugeot, AKKA Technologies, Vislab from
University of Parma, Oxford University, Google
■ these companies were more prevalent 2010-2015
DEMO I, II, and III
● US-funded military efforts
● demonstrated the ability of unmanned ground vehicles to navigate miles of
difficult off-road terrain
The Grand Challenges (I, II, and III)
● a fundamental problem in science or engineering, with broad applications,
whose solution would be enabled by the application of high performance
computing resources that could become available in the near future
● Grand Challenges were US policy terms set as goals in the late 1980s for
funding high-performance computing and communications research
DARPA Grand Challenge (2004)
● DARPA (Defense Advanced Research Projects Agency)
● March 13, 2004 in the Mojave Desert
● No cars finished
● Sandstorm from CMU traveled furthest: 11.78 km (7.32 mi)
Grand Challenge II (2005)
● 6:40am on October 8, 2005
Grand Challenge III (2007) aka Urban Challenge
● November 3, 2007 at the site of the now-closed George Air Force Base
● 96 km (60 mi) urban area course, to be completed in less than 6 hours
● obey all traffic regulations while negotiating with other traffic and obstacles
and merging into traffic
Google
Google’s Technology
● $150,000 in equipment including a $70,000 LIDAR system

● The range finder mounted on the top is a Velodyne 64-beam laser. This laser
allows the vehicle to generate a detailed 3D map of its environment.

● The car uses data collected from these mechanisms to drive itself.
Google’s Technology
How it works: Lidar system
● Laser + radar
● The system detects obstacles and tells the car when to avoid them to
navigate safely.
● It uses a 3D point cloud output provide the necessary data for robot software
to determine where potential obstacles exist in the environment and where
the car is is located relative to those obstacles.
How it works: Velodyne
● Company started experimenting with laser distance in 2005 with the DARPA
Grand Challenge
● Since then, they have vastly reduced the size of the sensor and weight while
improving its performance.
● It is a premier lidar system
How does communication among driverless cars
work?
● vehicles and roadside units as the communicating nodes

○ DSRC devices- 5.9 GHz band with bandwith of 75 MHz- range of 1000m
Communication among driverless cars cont.
● Smart intersections

○ intersections with no lights that communicate for autonomous cars

○ 2012- University of Texas in Austin

Google’s Track Record

● As of July 2015, Google’s cars have been involved in 14 “minor accidents”.

○ only one had resulted in minor injuries
● They’ve logged 1.7 million miles, and Google claims not a single collision was
caused by the self-driving mechanisms
Are we going to see Google on the road soon?
Google plans to make these cars available to the public in 2020.
Other Companies involved (since 1987)
Mercedes-Benz Audi

General Motors Volvo

Bosch Peugeot

Nissan Uber

Renault Google

Toyota Tesla
Mercedes Benz
Audi
Tesla’s Current Auto Pilot
Potential advantages
● being able to get things done while in traffic or on the road

● increase road capacity

● fewer traffic collisions. Experts estimate 300,000 lives can be saved per
decade

● higher speed limits

● reduction in traffic police

● removal of limitations on drivers — age and sobriety won’t be an issue

Potential obstacles
● Liability for damage

● Resistance by individuals to forfeit control of their cars

● Software reliability

● Implementation of legal framework and establishment of government

regulations for self-driving cars

● Drivers being inexperienced if situations arose requiring manual driving

● Loss of driving-related jobs

● Loss of privacy
Legislation
In the United States, state vehicle codes generally do not envisage — but do not
necessarily prohibit — highly automated vehicles.
Public Opinion
What do you think?

Would you be comfortable with an autonomous vehicle?

Public Opinion
● of 2,006 surveyed consumers, 49% would be comfortable
○ Accenture, 2011
● of 17,400 owners, 37% would be interested purchasing a self driving
○ 2012, J.D. Power and Associates
○ dropped to 20% if the technology costs $3000 or more
● of 1,000 German drivers, 10% undecided, 44% skeptical, 24% hostile
○ 2012, automotive researcher Puls
Discussion: Liability
● Situation: If a traditional automobile gets hit by a driverless car, who is
responsible?
● Opinion?
● Take a minute talk with the person next to you and decide what you think.
Discussion: Children
● Situation: Driverless cars may one day be able to pick a child up from school
and take him home if the laws permit
● Opinion?
Discussion: Licenses
● If driverless cars are a thing of the future, will driver licenses be a thing of the
past?
● Opinion?
● Take a minute talk with the person next to you and decide what you think.
●
Discussion: Morals
● If there was a choice to swerve into a schoolbus and potentially kill the
children onboard but save the driver, or divert the car to kill the driver but save
the children, how should the car be programmed?
● A real life application of The Trolley Problem
● Opinions?
● Take a minute talk with the person next to you and decide what you think.
●
Discussion: Jobs
● Will there still be a demand for auto insurance? What about public
transportation and taxi jobs, just to name a few?
● Opinion?
Predictions: Possible Developments
● By 2016, Mercedes plans to introduce "Autobahn Pilot" aka Highway Pilot, the
system allows a car to automatically pass someone while driving on a
highway.

● By early 2017, the US Department of Transportation hopes to publish a rule

mandating vehicle-to-vehicle (V2V) communication.

● By 2018, Elon Musk expects Tesla Motors to have developed mature serial
production version of fully self-driving cars, where the driver can fall asleep
behind the wheel.
Predictions: Possible Developments
● By 2018, Nissan anticipates to have a feature that can allow the vehicle
maneuver its way on multi-lane highways.

● By 2020, Volvo envisages having cars in which passengers would be immune

from injuries.

● By 2020, GM, Mercedes-Benz, Audi, Nissan, BMW, Renault, Tesla, Google

and Toyota all expect to sell vehicles that can drive themselves at least part of
the time

● By 2020, Google autonomous car project head's goal to have all outstanding
problems with the autonomous car be resolved.
SMART SPEAKER
CONSUMERADOPTION

REPORT
MARCH 2019
U.S.

G I V I N G V OI C E T O A REV OLU T I ON
Table of Contents About Voicebot About Voicify
Introduction // 3 Voicebot produces the l e adi ng online publication, Voicify i s the market leader in voice experience
newsletter and po dcas t fo cused on the voice and AI manage me nt software that co mbi ne s voice
Smart Speaker Ownership // 6 industries. Tho us ands of entrepreneurs, developers, optimized content manage me nt, cross-platform
investors, anal ys ts and other industry leaders look deployment, and voice-specific customer i ns i ghts.
Smart Speaker Use Cases //15 to Voicebot e ach week for the latest news, data, The Voicify Voice Experience Platform™ e nabl e s
anal ys i s and i ns i ghts de fining the trajectory of the marketers to connect with their cus to me rs by
Voice Assistants on Smart Phones // 22
next great co mputi ng platform. At Voicebot, we gi ve creating hi ghl y e ngagi ng and personalized voice
Voice App Discovery // 25 voice to a revolution. experiences that are automaticall y deployed to
a broad array of voice platforms s uch as voice
Consumer Sentiment about Smart Speakers // 29 as s i s tants (Amazon Alexa, Google As s i stant and
Microsoft Cortana), chatbo ts and other services.
Conclusion // 32 Methodology The platform e nabl e s non-technical us e r s to
deploy feature-rich voice appl i cati ons quickly and
Additional Resources // 33 The survey w as conducted online during the first
efficiently while offering the flexibility of unlimited
week of Januar y 2019 and w as completed by 1,038
customization.
U.S. adul ts age 18 or older that were representative
of U.S. C e ns us de mographic ave r ages . B e caus e we
reached only online adul ts which represent 89% of Voicify.com
the population acco r di ng to Pe w Re s earch Center,
s o me totals are adj us te d downward to provide
device and us age numbe r s relevant to the entire
adult population. Other fi ndi ngs are relative to device
ownership and do not require adj us tment.
SMART SPEAKER CONSUMERADOPTIONREPORT

One in Four U.S. Consumers Have

Access to a Smart Speaker Today
Smart speakers continued to be popular in 2018 keeping up a torrid pace of consumer adoption. In
January 2018, there were 47.3 million U.S. adults with a smart speaker and by the end of the year
that rose to 66.4 million. That means 26.2% of all U.S. adults have access to a smart speaker.

Moving Past Early Adopters

The number of sma rt sp ea kers per user a lso rose more th a n 10 % from 1.8 in 2 0 1 8 to 2.0 in 2 0 19. That
su g g ests there are a b ou t 133 million sma rt sp eakers in use in the U.S. today. However, the ex p a n sion
in sma rt speaker ownership h a s a lso brought in more c a su a l u sers. Whereas over 6 0 % of sma rt
speaker owners in Ja n u a ry 2 0 1 8 identified th emselves a s daily users, less th a n 5 0 % did so a year
later. And, the number of device owners that c la im to u se their sma rt sp eakers never or rarely doubled
to 2 6 %. That seems like a natural evolution of early adopters b ein g more frequent u sers th a n the early
majority u sers comin g afterward.

Reg a rd less, wh en more th a n one-in-four con su mers are u sin g a device a n d its voice a ssista n t, the
med ia , brands, g a me makers, service providers, independent developers, a n d even g ov ern ments are
sure to take notice. This recognition is p la y ing out with more voice a p p s published. The number of
Alexa skills rose by 2.2 times to nearly 60,000 in the U.S. a lon e in 2 018. During the sa me period Google
Actions grew a t a slightly fa ster rate of 2.5 times to over 4,000.

© VOICEBOT.AI - All Rights Reserved 2019 PAGE

215
SMART SPEAKER CONSUMERADOPTIONREPORT

A Different Smart Speaker Ecosystem, but the Same Leaders Smart Speakers are Solidly in the Early Majority Market
Voicebot reported in the fal l of 2018 that Phas e 1 of s mar t speaker adoption One w ay we can put the current state of s mar t speaker adoption in pers pe cti ve i s
w as over and we were entering Phas e 2. The s e co nd phas e i s characterized by to consider a s tandar d technolo gy adoption life cycle first de velope d i n the 1950’s
the influx of more cas ual us e r s but al s o by the introduction of new product form at Iowa S tate University and popularized in the 1990’s by Geoffrey Moore.
factors and new manufactur ers.
The model po s i ts that about 16% of the user population will be “innovators” and
The mo s t s i gnificant of these change s has been the e me r ge nce of s mar t “early adopters” followed by 34% that will be amo ng the “early majority.” With more
di s pl ays . When Amazo n w as the only manufactur ers of these voice-first de vi ce s than 26% population adoption, s mar t s pe ak er s are securely in the “early majority”
with di s pl ay screens, adoption w as mi ni mal. However, the introduction of Google s e gme nt today.
As s i s tant e nabl ed s mart di splays has helped drive s al e s , including Amazon, as it
An interesting as pe ct of mo vi ng al o ng the adoption curve i s that later adopters
brought more attention to the product category.
have different preferences than early adopters. Two ar e as of difference are
There are als o many more manufactur ers today than in 2017. B i g names in audio typically pl aci ng higher val ue in broader feature s e ts and integrations with other
s uch as Bose, B ang & Olufse n, and Klipsch all entered the s mar t speaker s e gme nt de vi ce s . You should expect to s e e s mar t speaker mak e r s e mphas i ze features,
in 2018 offering more co ns umer choice. However, the mo s t s igni fi cant new s mar t convenience of acce s s , and third-party integrations more in the co mi ng year.
speaker l aunch in 2018 w as Apple HomePod. That appe ar s to have captured
a s i gni ficant number of new s al e s in Q1 and Q2, but s e e ms to have tapered o ff
in Q3 and Q4. Although Apple w as threatening to break up the s mar t speaker 2019
duopoly, it appe ar s that Amazo n and Google enter 2019 nearly as strong as they
did in 2018 by mai ntai ni ng 85% in total installed bas e market share.

Innovators Early adopters Early majority Late majority Laggards

U.S. Smart SpeakerAdoption Curve

Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019 PAGE

216
Smart Speaker
Ownership
SMART SPEAKER CONSUMERADOPTIONREPORT

U.S. Smart Speaker Owners Rise 40% in 2018

The percent of U.S. adul ts that own s mar t s pe ak er s rose 40.3% w as under two ye ar s from the time when there w as more than
in 2018 cl i mbing from 47.3 million to 66.4 million during the one manufacture r participating in the category to reach that
year. This increase me ans more than one-in-four U.S. adul ts milestone.
now have acce s s to a s mar t speaker bas e d voice as s i s tant.
CES 2019 reaffirmed that including a s mar t voice as s i s tant i s no
We have mo ve d pas t the notion that s mar t s pe ake rs may be longer a differentiator for ho me speaker mak e r s . Like Bluetooth
a novelty as they are now in s uch widespread us e that one- before it, mak i ng s pe ak e rs s mar t has be co me a mus t-have
in-three s mar tphone owners have one. It took fewer than four feature. We now have do ze ns of s mar t speaker mo de l s from
ye ar s for s mar t s pe ake rs to achi e ve 25% adoption from the numerous manufacture s, but soon will have hundreds to cho o s e
initial introduction restricted to Amazo n Prime me mbe r s . And, it from even if 85% of us e r s favor just two device mak e r s .

U.S. Adult Smart Speaker Installed Base January 2019

Total US Adult Population 39.8%

253 MILLION One-Year Growth

Ja n 2019 / 66.4 MILLION

Ja n 2018 / 47.3 MILLION

Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019 PAGE 6
SMART SPEAKER CONSUMERADOPTIONREPORT

Amazon Echo Leads But Google Home Narrows the Gap

Amazo n continued to have the l e adi ng installed bas e of s mar t s pe ak ers in 2018 despite its market
s har e shrinking from about 72% to 61%. Google w as a bi g mover s hi fting from 18.4% to nearly 24%,
acco unti ng for precisely hal f of Amazon’s market s har e decline. U.S. Smart Speaker Market Share by Brand
January 2018 &2019
The “Other” category w as led by Apple and Sonos, and overall the non-Amazon, non-Google device
market s har e rose by 50% over 2017. More than hal f of this growth i s attributed to Apple HomePod
2019
which had a strong debut in the first hal f of 2018, but then tapered o ff in s al e s as the year went on.
There were several new s mar t s pe ak ers introduced in 2018 and many fo cus e d on s o und quality. It
61.1% 23.9% 15.0%
appe ar s co ns ume rs are open to addi ng these higher end s mar t s pe ak er s to their device collection as
Amazon Google Other
over three-quarters of “Other” category s mar t speaker owners al s o report havi ng either an Amazo n
Echo or Google Home device.
2018
S o no s went public in 2018 and w as clear in its investor do cume nts that voice as s i s tant integration
w as critical to the co mpany’s future competitive ness. However, the inability to l aunch a Google
As s i s tant e nabl ed speaker may have hurt its appe al with co ns ume r s as the co mpany’ s overall s mar t 71.9% 18.4% 9.7%
Amazon Google Other
speaker market s har e fell during the year. We can s ur mi se that mo s t of the S o no s fans that wanted
an Alexa-based speaker already bought their device in 2017. As the overall market e xpande d in 2018,
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
fe w additional S o no s One de vices were pur chased and the co mpany’ s relative market s har e fell.
Adding Google As s i s tant support in 2019 may help reverse this market s har e slide.

© VOICEBOT.AI - All Rights Reserved 2019

SMART SPEAKER CONSUMERADOPTIONREPORT

Low Price Speakers Account for 43% of Market

Amazo n Echo Dot i s the mo s t widely adopted s mar t speaker by a s i gni ficant
mar gi n. The s ub $50 list price device i s frequently avai l able for l e s s than $30 and
U.S. Smart Speaker Market Share by Device - January 2019
refurbished mo de l s can be acquired for under $20. This device has proven more
popular than Amazon’s higher priced o ffe r ings s uch as the Echo, Echo Pl us , Echo
31.4% 11.2% 10.0%
Spot, and Echo Show.
Amazon Echo Dot Google Home Other
In the Google portfolio, the Home and Home Mini appe ar to be equally popular
with 11.2% s har e e ach. There are likely to be more Home Mi ni s in us e today in
terms of total de vi ce s as this anal ys i s reflects the number of us e r s with acce s s to
a device. If you have one Home and three Minis, you are counted as one in e ach
category. And, this may be co mmo n as 87% of Google s mar t speaker owners
report havi ng both de vi ce s. 11.2%
23.2% Home Mini
Apple HomePod and S o no s One l e ad with s mar t speaker market s har e in the Echo or Plus
“Other” category. It appe ar s that s mar t di s pl ays with Google As s i s tant al ong with
2.7%
the introduction of Apple HomePod in February 2018 were the key drivers l e adi ng Apple
to a 50% growth in this category during the year. Keep in mi nd that as i de from HomePo
HomePod, the “Other” category de vi ce s all have Alexa or Google As s i stant on d

board, s o the do mi nance of Amazo n and Google voice as s i s tants extends beyond 3.5% // Echo Spot
2.2%
1.2% // Home Hub Sonos One
their own products. 3.0% Voicebot
Source: // Amazon Echo Show
Smart Speaker Consumer Adoption Report Jan 2019 0.2% // Home Max

© VOICEBOT.AI - All Rights Reserved 2019

SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Display Ownership Rises Quickly, Amazon Leads

The Amazo n Echo S ho w debuted as the first s mar t speaker
U.S. Smart Display Adoption by Smart Speaker Owners
with a di s pl ay screen now known as a s mar t dis pl ay in June
2017. Later in 2017, Amazo n als o l aunched the smalle r Echo
January 2019
Spot. With Amazo n as the singl e s mar t dis pl ay manufacturer, 13.2%
only 2. 8% of all s mar t speaker owners had adopte d one of the
de vi ce s in2017. September 2018
7.1%
This figure rose rapidly in early 2018 as Amazo n e ngage d in
aggr e s s i ve di scounting of the de vi ce s and then later in the year May 2018
5.9%
after manufacture rs starting introducing s mar t di s pl ays driven
by Google As s i stant. By year-end 2018, s mar t di s pl ays were
owned by 13.2% of s mar t speaker owners, a 558% growth rate January 2018
2.8%
in total installed bas e from about 1.3 million to 8.7 million.
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

U.S. Smart Display Market Share 2018

Today, s mar t di s pl ays in the U.S. are either Alexa or Google
As s i s tant e nabl ed r egardles s of the manufacturer. Although
33.0% 67.0%
Amazo n doesn’t have many third-party s mar t di s play OEM
Google Amazon
Assistant Alexa partners, it has mai ntai ned 67% market s har e in the category.
That me ans Google As s i stant e nable d de vi ces rose from zero
to one-third market s har e in l e s s than six mo nths . This may
have risen faster if Google’s s mar t display, the Home Hub, had
l aunche d earlier in the year. Despite not appe ar ing for s al e until
October 2018, Home Hub captured 38.5% of Google As s i s tant
s mar t di s play s al es .
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
© VOICEBOT.AI - All Rights Reserved 2019
SMART SPEAKER CONSUMERADOPTIONREPORT

U.S. Smart Speaker Frequency of Use 2018 New Smart Speaker Owners are Less
63.6%
Likely to be Daily Users
Maybe the bi gge s t change in the composition of s mar t speaker owners i s the
47.4% influx of more cas ual us e r s of the de vi ce s . Nearly 64% of device owners in
Januar y 2018 reported be i ng daily users. In Januar y 2019, that number fell to only
about 47%. Monthly us e r s were fairly similar with the offsetting difference be i ng
the infrequent us e r s which rose from 13% to over 26%.
26.5% 26.1%
23.5%
This s e e ms like a natural progression. Early
12.9% adopters of technology are more likely to
incorporate them quickly into their daily habi ts than
co ns ume rs that tend to adopt later. However, this
2018 2019 2018 2019 2018 2019 will be a metric to monitor go i ng forward. Three
NEVER ORRARELY MONTHLY DAILY
out of four s mar t speaker owners still report be i ng
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019 monthly active users. As long
as we s e e that type of consistent us age al o ng
with continued growth, s mar t s pe ak e rs will
continue to grow in importance as a voice
as s i s tant channel for co ns umer e ngage ment.

© VOICEBOT.AI - All Rights Reserved 2019

SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Speakers Per Household Rise to 2.0

Over 40% of s mar t speaker owners now have multiple de vi ce s. That i s up from 34% in 2018. Households
with two, three, and more than five de vi ce s all rose in 2019 as a percentage of all s mar t speaker owners. This
change s ugge s ts that many ho us e ho lds are fi nding utility in s mar t s pe ake rs be ing nearby.

The data indicate that the industry sold about 48 million s mar t s pe ak er s in the U.S. in 2018 bringing the total
in us e to about 133 million up from about 85 million at the end of 2017. Of the 19 million new s mar t speaker
owners, 31% have pur chase d multiple de vi ce s . That co mpar e s to 49% of U.S. adul ts that have owned s mar t
s pe ak e rs for more than a year and have multiple de vi ce s.

Smart Speakers Per Household -U.S.

0.7% / 5-10 devices
2.3% / 5-10 devices
1.4% / 10+ devices 3.2% / 4 devices 0.4% / 10+ devices
3.3% / 4 devices

8.0% 14.4%
3 devices
3 devices

65.7% 58.1%
19.3% 1 device 1 device
2 devices

23.2%
2 devices

2018 2019
© VOICEBOT.AI - All Rights Reserved 2019 Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019 PAGE 11
SMART SPEAKER CONSUMERADOPTIONREPORT

Living Room and Bedroom are Most Popular Locations

For the s e co nd straight year, the living room w as the Where Consumers Have Smart Speakers
mo s t co mmo n location for s mar t s pe ak ers . At just
under 45% it w as ahe ad of the bedroom at 37.6%
2.3% // Garage
which had about the s ame percentage as l as t year 37.6% 14.4%
but mo ve d up from third to the s e co nd mo s t popular Bedroom Home Office 32.7%
spot. Third pl ace went to the kitchen. At right around
Kitchen
33%, the kitchen s e e ms to have fallen from favor a
bit amo ng s mar t speaker owners. It’s still popular, 44.4%
but down from 41% in Januar y 2018. Living Room
Mo s t of the other locations were fairly similar to l as t
year with the exception of the ho me office which
grew by about one-third. As co ns ume rs have been
addi ng more s mar t s pe ak er s to their collection,
the ho me office s e e ms to be a co mmo n s e co nd
location. 2.0%
6.2% // Bathroom 6.5% // Dining Room Work
Office
Note: Multiple responses accepted, numbers total more than 100%
© VOICEBOT.AI - All Rights Reserved 2019 Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019 PAGE 12
SMART SPEAKER CONSUMERADOPTIONREPORT

Amazon Prime and Gmail Users More Likely to be Smart Speaker Owners
AMAZON P RIME GMAIL USE RS
It will surprise fe w people that Amazo n Prime me mbe r s are Gmail us e r s are al s o more likely than all us e r s to own a s mar t
50% more likely to own a s mar t speaker and more likely to speaker, in this cas e by about 31%. However, Gmail us e r s are
own an Echo branded device. Amazo n Echo s mar t s pe ak er s no more likely than all us e r s to own a Google Home device.
co mmand a 70% market s har e amo ng Prime members, but In fact, they are al mo s t exactly representative of all us e r s
surprisingly al s o adopt Google Home products at al mo s t a when it co me s to Amazon, Google, and third-party branded
22% rate. Non-Prime me mbe r s are more likely to adopt third s mar t s pe ake rs. Whereas a Prime me mbe r s hip and Gmail us e
party s mar t s pe ake rs made by manufacture rs other than s ugge s ts a bi as toward early technology adoption, only the
Amazo n or Google. Prime me mbe r ship s e e ms to materially influence consumer
choice of s mar t s peake rs.

Amazon Prime Member Smart SpeakerMarket Share

U.S. Smart Speaker Ownership Rates
U.S. 2019

All Smart SpeakerOwners All

U.S. 26.2%
23.9% Adults
61.1% 15.0%
Google
Amazon Echo Home Other
Gmail
Users 38.6%
Amazon Prime Members
Amazon
70.0% 22.0%
8.0%
Prime 45.3%
Google Members
Amazon Echo Other
Home Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019 PAGE 13

Smart Speaker
Use Cases
 SMART SPEAKER CONSUMERADOPTIONREPORT

Questions, Music, & Weather Still Reign Supreme

al w ays . When you look at monthly and daily use, it provides a far more accur ate
For the s e co nd straight year, as k i ng general que s tio ns i s the top indication of why co ns ume rs are us i ng the de vi ce s and in many cas e s why they
us e cas e mo s t co mmo nly tried by s mar t speaker owners. may be buyi ng a s e co nd or third s mar t speaker for the ho me . For example, nearly
However, it i s not the top us e cas e employed on a monthly or one-in-four s mar t speaker owners s ay they s e t an al ar m to pl ay on their s mar t
daily bas i s . That distinction go e s to listening to s tr e ami ng music speaker daily. That would s ugge s t a device location for the bedroom may be co me
s e r vi ces as it did in 2018. Third pl ace both ye ar s w as as k i ng increasingly important.
about the weather which i s followed by Timers and Alarms in the
fourth and fifth positions. Number six in 2019 w as listening to the The bi gge s t variance i s s mar t ho me control which i s ninth in terms of “ever tried”
radio. and fourth for “daily active use.” You mus t have a s mar t ho me device to us e
this feature s o that automaticall y e l i mi nates s o me people from trial. However,
You may have noticed that four of the top five us e cas e s are what controlling l i ghts or thermostats are already daily functions and if you have s mar t
are considered first-party services. That me ans they are provided ho me de vi ce s for these features, then s w i tchi ng your habi ts from s mar tphone
by the voice as s i s tant natively. Two of the top six us e cas e s app control to voice interaction i s a relatively e as y change . What s mar t speaker
involve mus i c which are third-party entertainment services. and voice as s i s tant developers want to s e e i s co ns umers us ing these de vi ce s
Positions 7-9 all go to the more traditional third-party services, frequently and incorporating them into daily routines. This not only l e ads to a
many of which were made by independent developers of Alexa higher perception of val ue by co ns ume rs but al s o l e ads to s ti ckiness which
skills and Google Actions. So, the order of us e frequency at a me ans the de vi ce s are l e s s likely to be removed or s w appe d out by co ns ume r s for
category level are first-party utilities, third-party entertainment, and a co mpe ti ng product.
third-party apps and services.

Frequency Sometimes
© VOICEBOT.AI - All Rights Reserved 2019 More Important Than Trial PAGE 15

You will notice that many anal ys e s of s mar t speaker us e only

fo cus on what us e r s have tried. This offers a pretty solid
gui de to what us e r s value, but not
SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Speaker Use Case Frequency January

2019
Ask a q u est ion 84.0% 66.0% 36.9%
List en to st rea min g mu sic serv ic e 83.0% 69.9% 38.
C h ec k t h e wea ther 80.1% 61.4% 35.6% 2%

Set a n a la rm 62.4% 41.8% 23.5%

Set a timer 62.4% 46.7% 22.9%
List en to ra d io 54.9% 40.5% 21.2%
U se a f a v orit e Alex a skill / Goog le Ac t ion 48.7% 35.0% 18.3%
P la y g a me or a n swer trivia 48.0% 29.1% 10.8%
Control sma rt h ome d ev ic es 45.8% 33.3% 23.5%
List en to n ews or sp ort s 43.8% 28.8% 13.4%
Sea rc h for p rod u c t inf o 41.2% 27.8% 10.8%

C a ll someon e 40.2% 23.5% 11.4%

Fin d a rec ip e / c ookin g in st ru c t ion s 40.2% 26.1% 7.8%
List en to p od c a st / other t a lk f orma t s 39.9% 26.5% 11.1%

C h ec k t ra f f ic 36.9% 22.9% 11.8%

Ac c ess my c a len d a r 31.7% 21.2% 11.4% EVER TRIED

Sen d a text messa g e 30.4% 18.3% 10.5% MONTHLY

M a ke a p u rc h a se 26.1% 15.0% 3.9% DAILY

© VOICEBOT.AI - All Rights Reserved 2019 Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019 PAGE 16
SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Home and Voice Commerce Rise

SMART HOME Monthly Active U.S.Smart Home Users on Smart Speakers
S mar t ho me us e cas e s were a notable mover between 2018-19.
In 2018 it w as the fourteenth mo s t co mmo n us e cas e and it
rose to ninth in 2019. Over 45% of s mar t speaker owners have
us e d them to control s mar t ho me de vi ces up from only 38% in
2018. And, one-third of s mar t speaker owners report now us i ng
voice for s mar t ho me device interactions on a monthly bas i s ,
29.9% 33.3%
up from 30% in 2017. At one time, it w as as s ume d that s mar t
2017 2018
speaker adoption w as be i ng driven by s mar t ho me afi cionado s.
It may be that the large audi e nce of s mar t speaker owners i s
now the key catal ys t for further s mar t ho me adoption.
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

VOICE COMMERCE
Voice co mme r ce w as a mover for a different reason. It had
Monthly Active U.S. Smart Speaker Voice Commerce Users
the lowest frequency of us e cas e s tracked this year. However,
it al s o s ho we d relative growth in monthly active us e r s during
2018. Monthly active us e r s rose 10.5% from 13.6% to 15.0%.
This i s still a relatively new us e cas e that co ns ume r s are
be co mi ng accus tome d to, but the growth i s indicative of the 13.6% 15.0%
utility of s ho ppi ng by voice. And, this isn’t just us e r s s e arching 2017 2018
for products. The r e s po nse s were s pe cific to mak i ng purchases .
When it co me s to product search, over 40% of us e r s have
attempted this us e cas e on s mar t s pe ak er s and 28% do s o
monthly. These are figures that are increasingly difficult for
co ns ume r br ands to ignore.
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
© VOICEBOT.AI - All Rights Reserved 2019 PAGE 17
SMART SPEAKER CONSUMERADOPTIONREPORT

Music, News & Movies Categories Lead in QuestionTopics

Question Category Frequency on Smart Speakers The mo s t tried s mar t speaker feature
and s e co nd hi ghe st frequency
monthly and daily us e cas e i s as k i ng
Music 54.9%
general que s ti ons. This feature
News 35.6%
turns a s mar t speaker into a voice
Movies 34.6%
interactive s e ar ch e ngine . Mus ic
How to instructions 28.4%
related que s tio ns are by far the mo s t
History 25.8%
co mmo n at about 55% of users. This
Products 24.8% w as followed by ne w s and mo vi e s
Restaurants 23.2% which were both topics identified
Sports 22.9% by over one-third of s mar t speaker
Retail store hours 22.6% owners.
Science 21.2%
Math 17.7% The next tier of que s ti ons clusters the
Games 16.7% 20-30% r ange of us e r s r angi ng from
Health and wellness as k i ng for how-to instructions and
16.0%
None of the above product information to retail store
16.0%
Celebrities hours and s ci e nce . Topics that are far
15.7%
Politics l e s s co mmo n include work-related
11.4%
information, fi nance and investing,
Local Business 10.1%
Travel
and fas hi o n. These were all registered
8.2%
by 5% or fewer s mar t speaker
Other 6.9%
owners.
Professional / work related topics 5.2%
Finance, banking, orinvesting 4.3%
Fashion 2.9% Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019

Reserved 2019 PAGE 18
SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Home Devices Popular with Smart Speaker Owners

S mar t ho me de vi ce s are popular amo ng Smart Home Devices Used by U.S. Smart Speaker Owners
s mar t speaker owners. More than 55%
of s mar t speaker owners s ay they have
at l e as t one s mar t ho me device that they
control by voice. Of course, be i ng abl e to
interact by voice with your s mar t ho me
de vi ce s doesn’t me an you are go i ng to
us e it as about one-in-five co ns ume rs 33.3% 21.2% 14.4% 12.4%
with s mar t ho me de vi ces have never tried
Smart TV Smart Lights S mart media controller, Smart Thermostat
controlling them with their s mar t speaker. game console or cable box

The mo s t popular s mar t ho me de vi ces by a

wide mar gi n are s mar t TVs with 33.3%. That
w as followed by s mar t lighting at 21.2%, 55.6%
voice interactive game co ns o l es and cabl e Have smart
boxes at 14.4% and s mar t thermostats at home devices 10.5% 8.8% 2.9% 2.0%
12.4%. Ro undi ng out the top five were video
doorbells at 10.5%. Video doorbell Smart cameras Smart appliances Other
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
© VOICEBOT.AI - All Rights Reserved 2019
SMART SPEAKER CONSUMERADOPTIONREPORT

Customer Service Coming to a Smart Speaker Near You?

A number of co mpani e s are starting to think about
how s mar t s pe ak er s can help addr e s s co mmo n
customer service que s ti ons. At a bas i c level, this could
be an FAQ from a co mpany’ s website repurposed U.S. Consumer Interest in Smart
for voice. The go al would be to provide customer Speaker Use for Customer Service
convenience and deflect inbound cal l center co ntacts
when practical. This could include s w i tching the
customer to a mobile, online, or cal l center channe l
30.4% 31.4%
when the voice app cannot ade quate ly addr es s the
inquiry.
Unsure Yes

A more advance d implementation mi ght include

acco unt linking and acce s s i ng information s pe ci fic
to a particular user. Finally, you could e nabl e use rs to
connect by voice to a live age nt to addr e s s detailed
i s s ue s that are be s t handl ed through a phone call.

Today, 31.4% of co ns ume rs would like to be abl e to

us e their s mar t s pe ake rs to contact customer service
departments. Another 30% are unsure and just under
40% are not interested. However, this me ans that more
38.2%
than 20 million U.S. adul ts are interested in do i ng this No
today and a co mpar abl e amo unt are open to it. This Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

could be a s i gni fi cant opportunity for customer service

departments to increase customer s ati s faction and
help them resolve i s s ue s more quickly.

© VOICEBOT.AI - All Rights Reserved 2019

Voice Assistants on Smart Phones
SMART SPEAKER CONSUMERADOPTIONREPORT

Voice Assistant Use on Smartphones Rising Quickly

Voice as s i s tant us e on s mar tphones rose quickly
during 2018, particularly in the s e co nd hal f of the Have Used a Voice Assistant on a Smartphone
year. While just under 56% of s mar tphone owners
had us e d a voice as s i s tant on their device in
Januar y 2018hat figure rose to over 70% one year
later. That me ans about 31 million U.S adul ts tried
us i ng a voice as s i s tant on their s mar t phone for the
first time in2018. 70.2%
J ANUARY
More than hal f of the growth w as driven by 2019
co ns ume rs trying Amazo n Alexa on their
smartphone, rising 124% during 2018. Before
Januar y 2018, us e r s could acce s s Alexa through
the s e ar ch bar in the Amazo n app. However, it w as
an obscure location for it and not well publicized.
That change d in early 2018 when an voice activation
button w as adde d to the Alexa app. These features
adde d to the actual Alexa app likely led to the s har p
increase in us e .

Bixby trial by co ns ume rs rose 57% and Google

56.7%
J ANUARY
As s i s tant just 16%, granted from a much larger
2018
bas e . The s mal l e st incremental gai n w as Apple Siri
which rose only 4% over all of 2018. It appe ar s that
Apple Siri, which has been around the longest, did
not have a discovery problem and that fe w us e r s
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
were introduced to it for the first time in 2018.
© VOICEBOT.AI - All Rights Reserved 2019
SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Speakers Increase Voice Assistant Use on Smartphones

S mar t speaker owners are 10% more likely to have us e d a voice as s i s tant on a Voice Assistant Use Frequency on Smartphones by Smart Speaker Ownership
smartphone. They are al s o more likely to be daily users. Of all voice as s i s tant
us e r s on s mar tphones 27.6% report be i ng daily users. Amo ng s mar t speaker
owners that figure rises to 39.8%. 49.3%
RARELY
One-third of s mar t speaker owners s ay after pur chas ing the device they are us i ng 26.8%
voice as s i s tants on their s mar tpho ne s more frequently. Jus t over 50% s ay they
are us i ng s mar tphone -bas ed voice as s i s tants about the s ame and only 14% s ay
A 31.5%
they are us i ng them l e s s . There i s a growing co ns e ns us that s mar t s pe ak er s are
T
di s pl aci ng time normally s pe nt on s mar tphones and many people posit that this L
will help reduce screen time. An accelerant for reducing screen time may be us i ng E
voice as s i s tants on s mar tphones as well. This reduces the touch, swipe, and look A
S 1
for many us e cas e s .
T
39.8%
S mar t speaker owners are about as likely as non-owners to be monthly voice
M 33.3%
O
as s i s tant us e rs on s mar tphones . However, they are twice as likely to be daily N Smartphone Voice Assistant Smartphone VoiceAssistant
T Use Frequency of Non Smart Use Frequency of Smart
users. Data i s consistently s ho w i ng that us age of voice on one platform i ncr e as e s H Speaker Owners Speaker Owners
us age of voice on others platforms. L
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
Y

© VOICEBOT.AI - All Rights Reserved 2019

A 9.1%
T
L
E
A
S
T
Voice App Discovery
SMART SPEAKER CONSUMERADOPTIONREPORT

Friends are the Most Common Source of Voice App Discovery

For the s e co nd year in a row, just under hal f of s mar t speaker owners s ay they How Smart Speaker Owners Discover Voice Apps
don’t actual l y discover new voice apps . One-in-four device owners rely on friends I don’t
to introduce them to new voice apps followed by just 15% that note s o ci al me dia 49.7%
as a discovery channel. Friends

A smaller group of us e r s between 11-14% cite Amazo n and Google’s primary 26.8%
promotion channe l s as key s o ur ce s of discovery, s uch as their in-app and online Social media
stores and weekly e mai l s . Not far behind these channe l s i s advertising which w as 15.4%
l e s s visible in previous years, but now i s a source of voice app discovery for one- Alexa skill store / Google Assistant discover section
in-ten s mar t speaker owners. 13.7%
Discovery i s the top i s s ue faci ng third-party voice app publishers today. The voice Email newsletter from Alexa or GoogleAssistant

as s i s tant user bas e i s growing quickly, but about hal f of these us e r s are only 11.1%
discovering first-party solutions provided by the voice as s i s tants the mse lve s Ads / commercials
s uch as Alexa and Google As s i s tant. Many third-parties are havi ng more trouble 10.5%
capturing new users. Word-of-mouth appe ar s to be the mo s t effective channel, News media
but i s the hardest to tap into. So, mo s t voice app publishers should fo cus on a 7.2%
variety of techniques r angi ng from ne w s me di a coverage and s o ci al me dia to Other
advertising to drive discovery today.
2.9%
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
© VOICEBOT.AI - All Rights Reserved 2019 PAGE 25
SMART SPEAKER CONSUMERADOPTIONREPORT

Number of 3rd Party Voice Apps Tried on Smart Speakers

The s ame ratio of s mar t speaker owners report hav- What do e s this tell us ? Well, there r e mai ns a dis-
i ng us e d third-party voice apps (i.e. those not made covery problem where about hal f of s mar t speaker
by the voice as s i s tant provider) at the end of 2018 owners have never tried a third-party voice app and
as the previous year. More than hal f of co ns ume rs relegate their us age to first-party voice as s i s tant
that try third-party voice apps have tried at l e as t interactions. However, once s mar t speaker owners
three and nearly 12% have us e d more than six. do employ a third-party voice as s i s tant, they are very
likely to be co me monthly users.
Even more interesting i s about 93% of s mar t s pe ak-
er owners that try third-party voice apps s ay they This i s go o d ne w s for third-party voice app publish-
be co me monthly users. And, 55% of these monthly ers that can drive discovery. If they can induce trial,
us e r s are acce s s i ng at l e as t two voice apps from they have a go o d chance of converting the introduc-
third parties and 35.5% are us i ng three or more. tion to regular us e .

Percent of Smart Speaker

Users That Try 3rd Party
Voice Apps and Convert to
Consumers That 51.3% 48.7% 92.6% Monthly Users
Have Have Become
Have Used 3rd Not Used Used Monthly
Party Voice Apps Users
3.6%
>10 1.3%
6 -10
6.6%
8.3% >10
6 -10
31.0% Number of 3rd
1 27.6% 44.7% Party Voice Apps
Number of 3rd Party 3 -5 1 Accessed by
40.5%
Voice Apps Tried on 3 -5 Monthly Users
Smart Speakers 16.7%
2 19.7%
2

Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019
SMART SPEAKER CONSUMERADOPTIONREPORT

Voice App Reviews Rise in 2018

Only 14% of s mar t speaker owners have ever left a review of a third-party voice U.S. Smart Speaker Owners That Have Left a Voice App Review
app. This i s up from just 11% at the end of 2017 and i s a reflection of the fact that
mo s t reviews mus t be submitted through a vi s ual interface for a user experience Once
de s i gne d for no vi s ual interaction. 6.9% More than once
7.2%
Amazo n introduced voice r ati ngs in late 2018 which e nabl e d us ers to offer a star 85.9%
rating for an Alexa skill by voice after us i ng it. However, this w as limited to a fe w Never
skills and not avai l abl e for skill publishers to s e t as a feature on their own. By
contrast, Google As s i s tant us er s are more likely to us e the voice as s i s tant both on
s mar tphones and s mar t s pe ak er s. That multimodal us e profile mi ght explain why
Google Home owners are about 11% more likely to have left a review than those
with Amazo n Echo de vi ces.

With all of that s ai d, 14% s e e ms like a s mal l number of s mar t speaker owners
l e avi ng reviews until you consider the fact that only 48.7% s ay they have even
us e d a third-party voice app. That me ans 29% of device owners that have tried a
third-party voice app have left a review. This i s a promising figure gi ve n the friction
involved in actual l y l e avi ng a review provided voice app publishers can increase
the proportion of s mar t speaker owners that try third-party apps .
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019
© VOICEBOT.AI - All Rights Reserved 2019
 Consumer Sentiment
About Smart Speakers
SMART SPEAKER CONSUMERADOPTIONREPORT

Consumers Want to Be Understood What Qualities U.S. Users Value in Smart Speakers
How well it understands
Two-thirds of co ns ume rs s ay how well a voice as s i s tant me when I speak 67.0%
unde r s tands them i s an important quality. That i s followed by
Sound quality 54.9%
s o und quality at 54.9%, how much it can do at 50.7%, and the
s pe e d of response at 45.1%. Every other quality i s well behind
the s e top four char acter istics. How much it can do 50.7%

Amazo n, Apple, and Google executives have s po k en many ti me s How fa st it responds 45.1%
abo ut their fo cus on addi ng personality to voice as s i s tants de s pi te
the fact that it i s considered important by only 15. 4% o f s mar t Its personality 15.4%
s pe ak er owners. That lower rating may be i nfl uenced by the fact
that personality i s offered by all of the l e adi ng voice as s i s tant Whether it has my
favorite media entertainment 14.4%
providers, but it i s clearly not s o me thing havi ng an i mpact today.
Whether the voice assistant is
the same as my mobile device 10.1%
It i s al s o notable that s mar tphone ownership only influenced
s mar t speaker selection for about one-in-ten co ns ume rs. Apple I am not interested
in a smart speaker 9.8%
and Google would like that linkage to be higher gi ve n their
do mi nance of s mar tpho ne-base d voice as s i s tants worldwide.
Whether it has goodgames 4.6%
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019 PAGE 29

29
SMART SPEAKER CONSUMERADOPTIONREPORT

Privacy Concerns Don’t Preclude Smart Speaker Ownership

There i s a lot of talk in the me di a about consumer

privacy fe ar s related to havi ng a s mar t speaker U.S. Consumer Perception of
in the household. Those fe ar s don’t s e e m to be
Smart Speaker Privacy Risk
undermining adoption, but it i s true that two-thirds
of co ns ume r s express at l e as t s o me privacy
concerns and 26% are very concerned.

Interestingly, the privacy concerns for all co ns ume r s

11.8%
I don’tknow
and those that do not own s mar t s pe ak e rs are 21.5%
nearly identical. For example, only 27.7% of Notconcerned
co ns ume rs without s mar t s pe aker s s ai d they were
very concerned about privacy i s s ue s compared to
21.9% of device owners. This me ans that even s o me
26.0%
Very
co ns ume rs with privacy concerns went ahe ad and
pur chas ed s mart s peake rs. Those co ns ume rs very
concerned 20.8%
Mildy
concerned about privacy risks are about 16% l e s s concerned
likely to own a s mar t speaker, but they still mak e up
a s i ze able proportion of users. 20.0%
Moderately
Apple has made a bi g de al about as s e r ting that Siri concerned
and the iOS e co s ys te m are more protective of user
privacy than its competitors. However, that doesn’t Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

s e e m to be mak i ng much impression on co ns ume rs

as the mo s t privacy concerned co ns ume rs are
actual l y slightly l e s s likely to own a HomePod.

© VOICEBOT.AI - All Rights Reserved 2019 PAGE 30

SMART SPEAKER CONSUMERADOPTIONREPORT

About 10% of Consumers Without Smart Speakers are Interested

Only about one-third of co ns ume rs without s mar t Non Smart Speaker Owners Opinions AboutDevices
s pe ak e rs have no interest in the product category. That
i s down from 38% at the end of 2017. The number of
No opinion / not interested
non-owners that s ay they don’t need s mar t s pe ake rs
be caus e of the capabilities of s mar tpho ne s i s up from 34.2%
21.2% to 24.2%. My smartphone has all the functionality I need

Also, the number of co ns ume rs without de vi ce s

29.6%
that expect to pur chas e one fell from 11.8% to 9.8%. Not interested
However, s e ve n out of 10 of those prospective s mar t 24.2%
speaker owners s ay they will acquire their first device Concern that the device will record what I'm saying
this year. That would lift total s mar t speaker ownership
to about one-third of U.S. adul ts in 2019.
23.0%
They’re too expensive
Another interesting data point i s that 12.2% of
12.2%
co ns ume rs without s mar t s peake rs s ai d this year that
I hope to get one this year
price w as a deterrent for acqui r ing a s mar t speaker
up from 8.8% l as t year. That i s surprising considering 7.1%
that discounted prices for s mar t s pe ake rs were more I hope to get one after this year
aggr e s s i ve in 2018 than in 2017. With that s ai d, there
2.7%
were many new s mar t s pe ak er s in the premium
Waiting for Samsung Galaxy Home
category that stretched the pricing umbrella upw ar ds
despite more options at lower prices. 1.0%
Source: Voicebot Smart Speaker Consumer Adoption Report Jan 2019

© VOICEBOT.AI - All Rights Reserved 2019 PAGE 31

31
SMART SPEAKER CONSUMERADOPTIONREPORT

Smart Speakers Are a Catalyst for Voice Adoption

The trend line for s mar t speaker adoption r e mai ns on an upward A Catalyst for Smart Home Adoption
slope and has quickly s ur pas s e d 25% of the U.S. adult population.
More than three out of four co ns ume rs that didn’t own a s mar t It i s widely believed that the first w ave of s mar t speaker adop-
speaker but pl anne d to buy one in 2018 did s o . With that s ame tion w as driven in part by s mar t ho me early adopters. It w as one
follow-through rate, more than one-third of U.S. adul ts will own a of the first robust us e cas e s for the de vi ce s . However, it may be
s mar t speaker at this time next year. That will reflect 33% penetra- that the catal ys t behind the next w ave of s mar t ho me adoption
tion in five ye ar s from the first product l aunch in the category and will be the growing s mar t speaker user bas e . If you exclude
just three ye ar s s i nce there were two vendors offering solutions. s mar t TVs and voice-interactive me di a boxes and fo cus just on
ho me automation, there are now more s mar t speaker house-
Those first two vendors, Amazo n and Google, continue to dom- holds. That me ans s mar t speaker ownership can be a fertile
inate the market with 85% market share. That i s down 5% from source of new cus to mer s for s mar t ho me automation vendors.
the previous year, but the s har e i s even higher when you consider
ho us e ho lds with multiple de vi ce s. When it co me s to s mar t s peak- The Battle for New Use Cases and Features
er us e r s in the U.S., Amazo n and Google have created a duopoly
S mar t speaker competition thus far in the U.S. has been char-
that i s likely to l as t for s o me time.
acterized by price competition and hardware features s uch as
A Catalyst for Voice Adoption Across All Surfaces s o und quality and di s pl ay availability. Over the next year, s mar t
speaker competition of third-party manufacturer s will continue
S mar t s pe ak er s may have already done their job promoting voice to fo cus on these ar e as . However, Amazon and Google will s e e k
as s i s tants as one-third of device owners report us i ng them on to differentiate the ms e lve s more on new features and us e cas e s
s mar tphones more frequently after pur chas e. We are s e e i ng voice e nabl e d by their respective voice as s i s tants that increase the
as s i s tants through s mar t s pe ak er s be come more deeply embed- perceived daily val ue of the de vi ce s . Voicebot reported in the
ded in household routines while al s o i ncr easi ng us e on other sur- fal l of 2018 that we have entered phas e two of voice as s i s tant
face s s uch as s mar tphones and automobile das hbo ar ds . S mar t adoption which will be characterized by a fo cus on daily habit
s pe ak e rs will continue to be an important co ns ume r touchpoint formation and broader us age by the installed bas e w he r e as
and likely the catal ys t for increased voice as s i s tant us e on other phas e one w as more about device adoption. That trend will
de vi ce s beyond co mmuni cati ons , navigation, and alerts. continue in 2019.
© VOICEBOT.AI - All Rights Reserved 2019