we trained a neural network on satellite imagery to predict wealth from space
TRANSCRIPT
Quartiles
The Training Model
Quartile 0< $34,176
Quartile 1$34,177 - $49,904
Quartile 2$49,905 - $71,875
Quartile 3$71,876+
Satellite Imagery & Census Data Output
=
Neural Network
The Training Model (Resnet 50)
Q0 = 91%Q1 = 5.64%Q2 = 2.55%Q3 = .41%
Satellite Imagery & Census Data Output
=
Neural Network
The Training Model (Resnet 50)
The model carrieswhat it has learned and
repeats the process.
Q0 = 91%Q1 = 5.64%Q2 = 2.55%Q3 = .41%
The Training Model (Resnet 50)
Input Normalization Normalization Normalization MergeActivationActivationConvolution Convolution Convolution
Neural Network
The data is fed into the model.
Input Normalization Normalization Normalization MergeActivationActivationConvolution Convolution Convolution
The Training Model (Resnet 50)Neural Network
A number of filters are applied to the image.
Input Normalization Normalization Normalization MergeActivationActivationConvolution Convolution Convolution
The Training Model (Resnet 50)Neural Network
The resulting new values are normalized to be within learned mean and standard deviations of the dataset.
Input Normalization Normalization Normalization MergeActivationActivationConvolution Convolution Convolution
The Training Model (Resnet 50)Neural Network
Separates out features from important and non important ones.
Input Normalization Normalization Normalization MergeActivationActivationConvolution Convolution Convolution
The Training Model (Resnet 50)Neural Network
Repeat.
Input Normalization Normalization Normalization MergeActivationActivationConvolution Convolution Convolution
The Training Model (Resnet 50)Neural Network
Results are merged with previous iteration.
“Despite this encouraging process, there is still little insight into the internal operation and behavior of these complex models, or how they achieve such good performance. From a scientific standpoint, this is deeply unsatisfactory. Without clear understanding of how and why they work, the development of better models is reduced to trial-and-error.”
Visualizing and Understanding Convolutional Networks - Matthew D. Zeiler, Dept. of Computer Science, Courant Institute, New York University - Rob Fergus, Dept. of Computer Science, Courant Institute, New York University
Baseball Field Experiment
Confidence 63% Q0 (-28)19% Q1 (+14)13% Q2 (+11)3% Q3 (+2.6)
Added More Trees
Baseball Field Experiment
Confidence 68% Q0 (-23)17% Q1 (+12)11% Q2 (+9)2% Q3 (+1.6)
And Added More Trees
Baseball Field Experiment
Confidence 44% Q0 (-47)12% Q1 (+7)23% Q2 (+21)18% Q3 (+17.6)
Added All The Trees
Conclusions:• It's possible to predict income levels
from space• Underlying data can provide valuable
assistance to complex neural networks• Human-based empirical inquiry has legs• Teasing out why it knows what it knows
is interesting
Questions:• What does this thing do when you point
it at other cities?• What are the similarities and differences
between cities from space?• Can we construct a model to account for
seasonal variance?• Can we construct a model to account for
architectural difference?
cmu.edu
Who
stamen.com amantiwari.comgbdx.geobigdata.io
Carnegie MellonStamen Design Aman TiwariDigital Globe