Ishanu Chattopadhyay
Assistant Professor of
Data Science
University of Chicago
Rare and Extreme Events In Complex Dynamical Systems
Urban Crime
Extreme weather phenomena
Seismic events
Long-range memory
Non-trivial stochastic effects
Can we learn stochastic phenomena non-parametrically?
Can we learn stochastic phenomena non-parametrically?
?
State structure :
self-similarity in dynamical systems
The Fractalnet architecture
No back-propagation
Deep Learning without Neural Networks
Lets Predict Crime
Philadelphia
The Problem:
>3 days in advance
Within ~2 city blocks
ONLY Past eventlog as input
Mean AUC
Property crime: 81%
Violent crime: 84%
Spatial tiles:
0.003 deg latitude, 0.003 deg longitude
0.25 miles across
Time-period:
Training: Jan 1 2016 - Dec 31 2018
Out-of-sample test: Jan 1 2019 - April 1 2019
Prediction Performance (Philadelphia)
sensitivity 0.90
ppv 0.87
100 crimes
Raise 103 flags
90 correct flags
13 false positives
10 missed
3 day ahead prediction
Jan 1 2019
to
April 1 2019
Play Movie
Triangles: actual events
heatmap: predicted risk 3 days ahead
Triple homicide incident
Jan 7 2019
https://www.inquirer.com/crime/kensington-triple-shooting-homicide-philadelphia-police-20190107.html
Triangles: actual events
heatmap: predicted risk 3 days ahead
Predicting
extreme weather
Outperforms pure physics-based models at longer horizons
Next:
Model and predict world events
Nearly
Hands-free
No feature-engineering
The Issue with Neural Networks
only certain specific structures may be identified