Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The...
Transcript of Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The...
![Page 1: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/1.jpg)
Interpreting and Explaining Deep Neural Networks:
A Perspective on Time Series Data – Part 3/3
Jaesik Choi
Explainable Artificial Intelligence Center
Graduate School of Artificial Intelligence
KAIST
Some slides courtesy of Eamonn Keogh
![Page 2: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/2.jpg)
Overview to Explainable Artificial Intelligence (XAI) – 15 min
Input Attributions Methods for Deep Neural Networks – 35 min
Interpreting Inside of Deep Neural Networks – 50 min
Explainable Models for Time Series Data – 50 min
- Important Questions on Mining and Learning Time Series Data
- Visualizing Deep Temporal Neural Networks
- Clustered Pattern of Highly Activated Period (CPHAP)
- Automatic Statistician/Relational Automatic Statistician (Bayesian Approaches)
Agenda (150 min)
Interpreting and Explaining Deep Neural Networks:
A Perspective on Time Series Data
![Page 3: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/3.jpg)
100 Time Series Data Mining Questions
Slides courtesy of Eamonn Keogh
![Page 4: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/4.jpg)
Time Series Motif
Slides courtesy of Eamonn Keogh
![Page 5: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/5.jpg)
Time Series Motif – An Example: Insect Behavior Analysis
Slides courtesy of Eamonn Keogh
![Page 6: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/6.jpg)
Time Series Motif – An Example: Insect Behavior Analysis
Slides courtesy of Eamonn Keogh
More motifs reveal different feeding patterns of Beet Leafhopper.
![Page 7: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/7.jpg)
Are there any repeated patterns in my data? – Motif Search
Slides courtesy of Eamonn Keogh
The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100 Hz.
Note that there may be more examples of each motif!
![Page 8: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/8.jpg)
Motif in Deep Temporal Neural Networks
Here, we are interested in finding Temporal MotifsTrained in Deep Temporal Neural Networks
![Page 9: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/9.jpg)
Temporal Neural Networks: MLP vs FCN vs ResNet
Wang, Z. et al., Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline, arXiv:1611.06455, 2016.
![Page 10: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/10.jpg)
Experimental Results on UCR dataset
Wang, Z. et al., Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline, arXiv:1611.06455, 2016.
![Page 11: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/11.jpg)
Experimental Results
https://github.com/cauchyturing/UCR_Time_Series_Classification_Deep_Learning_Baseline
Wang, Z. et al., Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline, arXiv:1611.06455, 2016.
![Page 12: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/12.jpg)
Residual Network [ResNet, He et. al., 2015]
Residual learning
Comparison of Resnet
3.6% of error in ImageNet Challenge, 2015
He et. al., 2015
![Page 13: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/13.jpg)
Recurrent Convolutional Neural Layers [RCNN, Liang and Hu, 2015]
Recurrent Convolutional Layer (RCL)
Liang and Hu, 2015
![Page 14: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/14.jpg)
Hand Start
First Digit Touch
Lift off
Replace
Both Released
* Joint work with Azamatbek Akhmedov
RCNN on EEG AnalysisLuciw et. al., 2014
![Page 15: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/15.jpg)
One chunk: Data: 3584,32
Hand Start
First Digit Touch
Lift off
Replace
Both Released
* Joint work with Azamatbek Akhmedov
RCNN on EEG Analysis
![Page 16: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/16.jpg)
Applying RCL Convolutional Layer:(1,3584)
Max pooling
Max pooling
Max pooling
Max pooling
Max pooling
Fully Connected
RCL:(1,896)
RCL:(1,224)
RCL:(1,56)
RCL:(1,14)
(1,7)
(6)
97.687%
RCNN on EEG Analysis
![Page 17: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/17.jpg)
RCNN on EEG Analysis
256 1x9 filters
![Page 18: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/18.jpg)
Example: Hand Start
RCNN on EEG Analysis
![Page 19: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/19.jpg)
Example: Hand Start
RCNN on EEG Analysis
![Page 20: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/20.jpg)
Example: Hand Start
RCNN on EEG Analysis
![Page 21: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/21.jpg)
Example: Hand Start
RCNN on EEG Analysis
![Page 22: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/22.jpg)
Example: First Digit Touch
RCNN on EEG Analysis
![Page 23: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/23.jpg)
Example: First Digit Touch
RCNN on EEG Analysis
![Page 24: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/24.jpg)
Example: First Digit Touch
RCNN on EEG Analysis
![Page 25: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/25.jpg)
Example: First Digit Touch
RCNN on EEG Analysis
![Page 26: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/26.jpg)
Example: Replace
RCNN on EEG Analysis
![Page 27: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/27.jpg)
Example: Replace
RCNN on EEG Analysis
![Page 28: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/28.jpg)
Example: Replace
RCNN on EEG Analysis
![Page 29: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/29.jpg)
RCNN on EEG Analysis
Example: Replace
![Page 30: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/30.jpg)
Image dataset
● How can we separate time series data into semi-global representative
parts without hand-crafted segmentation labels for interpreting?
Time Series dataset
?
?
● Network Dissection ● LRP
Clustered Pattern of Highly Activated Period: MotivationS. Cho et al., 2020
![Page 31: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/31.jpg)
CPHAPClassification
Clusters from Layer 1
Clusters from Layer 2
Clusters from Layer 3
Inp
ut
Clustered Pattern of Highly Activated PeriodS. Cho et al., 2020
![Page 32: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/32.jpg)
ClassificationIn
pu
t
Highly Activated Period (HAP)
Clustered Pattern of Highly Activated PeriodS. Cho et al., 2020
![Page 33: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/33.jpg)
Classification
Clusters from Layer 1
Clusters from Layer 2
Clusters from Layer 3In
pu
t
Clustered Pattern of Highly Activated PeriodS. Cho et al., 2020
![Page 34: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/34.jpg)
CPHAPClassification
Clusters from Layer 1
Clusters from Layer 2
Clusters from Layer 3
Inp
ut
Clustered Pattern of Highly Activated PeriodS. Cho et al., 2020
![Page 35: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/35.jpg)
Input
SequenceHAP CPHAP CPHAP with
Uncertainty
Clustered Pattern of Highly Activated PeriodS. Cho et al., 2020
![Page 36: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/36.jpg)
Layer 3 | Chanel 57 Layer 3 | Chanel 67 Layer 3 | Chanel 36
Layer 1 | Chanel 18 Layer 1 | Chanel 30 Layer 1 | Chanel 5
Layer 2 | Chanel 36 Layer 2 | Chanel 56 Layer 2 | Chanel 48
Clustered Pattern of Highly Activated Period: ResultsS. Cho et al., 2020
![Page 37: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/37.jpg)
Layer 3 | Chanel 57 Layer 3 | Chanel 67 Layer 3 | Chanel 36
Layer 1 | Chanel 18 Layer 1 | Chanel 30 Layer 1 | Chanel 5
Layer 2 | Chanel 36 Layer 2 | Chanel 56 Layer 2 | Chanel 48
Clustered Pattern of Highly Activated Period: ResultsS. Cho et al., 2020
![Page 38: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/38.jpg)
Layer 3 | Chanel 57 Layer 3 | Chanel 67 Layer 3 | Chanel 36
Layer 1 | Chanel 18 Layer 1 | Chanel 30 Layer 1 | Chanel 5
Layer 2 | Chanel 36 Layer 2 | Chanel 56 Layer 2 | Chanel 48
Clustered Pattern of Highly Activated Period: ResultsS. Cho et al., 2020
![Page 39: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/39.jpg)
Layer 3 | Chanel 57 Layer 3 | Chanel 67 Layer 3 | Chanel 36
Layer 1 | Chanel 18 Layer 1 | Chanel 30 Layer 1 | Chanel 5
Layer 2 | Chanel 36 Layer 2 | Chanel 56 Layer 2 | Chanel 48
Clustered Pattern of Highly Activated Period: ResultsS. Cho et al., 2020
![Page 40: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/40.jpg)
Layer 3 | Chanel 57 Layer 3 | Chanel 67 Layer 3 | Chanel 36
Layer 1 | Chanel 18 Layer 1 | Chanel 30 Layer 1 | Chanel 5
Layer 2 | Chanel 36 Layer 2 | Chanel 56 Layer 2 | Chanel 48
Clustered Pattern of Highly Activated Period: ResultsS. Cho et al., 2020
![Page 41: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/41.jpg)
Experiment 1 : Different Network Structure (ResNet)S. Cho et al., 2020
![Page 42: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/42.jpg)
Experiment 2 : Different Filter SizeS. Cho et al., 2020
![Page 43: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/43.jpg)
Experiment 3 : Sequences of test data with CPHAP of train dataS. Cho et al., 2020
![Page 44: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/44.jpg)
Experiment 4 : Visual Comparison among XAI methodsS. Cho et al., 2020
![Page 45: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/45.jpg)
Experiment 5 : Perturbating with unimportant area
https://clusteredpattern.github.io/pages/
S. Cho et al., 2020
![Page 46: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/46.jpg)
Automation of Knowledge Work
SOURCE: https://public.tableau.com/profile/mckinsey.analytics#!/vizhome/AutomationBySector/WhereMachinesCanReplaceHumans
Finance and Insurance, McKinsey 2016
![Page 47: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/47.jpg)
Automated Narrative Generation
Adobe beats Street 3Q forecastsAssociated Press September 20, 2017
SAN JOSE, Calif. (AP) _ Adobe Systems Inc. (ADBE) on Tuesday reported fiscal
third-quarter profit of $419.6 million.
The San Jose, California-based company said it had profit of 84 cents per share.
Earnings, adjusted for one-time gains and costs, were $1.10 per share.
…
Adobe shares have climbed 52 percent since the beginning of the year. In the
final minutes of trading on Tuesday, shares hit $156.61, an increase of 57
percent in the last 12 months.
_____
This story was generated by Automated Insights
(http://automatedinsights.com/ap) using data from Zacks Investment Research
Automated Insights, 2017
![Page 48: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/48.jpg)
Automated Narrative Generation
Sonoma County Little Leagues (Falcons vs Mustangs)
Anthony T got it done on the bump on the way to a win. He
allowed two runs over 2-1/3 innings. He struck out four, walked
two, and surrendered no hits.
Anders Mathison ended up on wrong side of the pitching
decision, charged with the loss. He lasted just two innings, walked
two, struck out one, and allowed four runs.
Automated generated by Quill, Narrative Science
Narrative Science, 2015
![Page 49: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/49.jpg)
Turing Test? Software vs Journalist
Each of 45 respondents read a game recap article and decide
whether or not the text had been written by a journalist or by a
computer.
Clerwall, Journalism Practice, 2014
![Page 50: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/50.jpg)
Big Success in Funding
Automated Insights is acquired by Vista for
$80 million (Feb. 2015).
Narrative Science get funded $43.4 million,
so far.
…
![Page 51: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/51.jpg)
An Old-School AI Strategy: Template
Adobe beats Street 3Q forecastsAssociated Press September 20, 2017
SAN JOSE, Calif. (AP) _ Adobe Systems Inc. (ADBE) on Tuesday reported fiscal
third-quarter profit of $419.6 million.
The San Jose, California-based company said it had profit of 84 cents per share.
Earnings, adjusted for one-time gains and costs, were $1.10 per share.
…
Adobe shares have climbed 52 percent since the beginning of the year. In the
final minutes of trading on Tuesday, shares hit $156.61, an increase of 57
percent in the last 12 months.
_____
This story was generated by Automated Insights
(http://automatedinsights.com/ap) using data from Zacks Investment Research
Automated Insights, 2017
![Page 52: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/52.jpg)
An Old-School AI Strategy: Template
Sonoma County Little Leagues (Falcons vs Mustangs)
Anthony T got it done on the bump on the way to a win. He
allowed two runs over 2-1/3 innings. He struck out four, walked
two, and surrendered no hits.
Anders Mathison ended up on wrong side of the pitching
decision, charged with the loss. He lasted just two innings, walked
two, struck out one, and allowed four runs.
Generated by Quill, Narrative Science
Narrative Science, 2015
![Page 53: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/53.jpg)
Challenges in Algorithmic Authors
The deeper challenge lies not in generating copy, but
in finding the most pertinent meaning in a given
dataset.
“It’s not just about converting numbers to language.”
“Those numbers need context”
Wright et. al., 2015
![Page 54: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/54.jpg)
Finding Context in Time Series Data
![Page 55: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/55.jpg)
Descriptive prediction of time series
Problem
![Page 56: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/56.jpg)
Descriptive prediction of time series
Problem
![Page 57: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/57.jpg)
Linear function
decrease x/week
Smooth function
Length scale: y weeks
Rapidly varying
smooth function
Length scale: z hours
Descriptive prediction of time series
Problem
![Page 58: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/58.jpg)
𝑓 𝑥 ~𝒢𝒫(𝜇 𝑥 , 𝑘 𝑥, 𝑥′ )
Function Gaussian Process
Mean function
Covariance kernel function
𝜇 𝑥 = 𝔼 𝑓 𝑥
𝑘 𝑥, 𝑥′ = Cov 𝑓 𝑥 , 𝑓 𝑥′
Gaussian Processes (GP)
![Page 59: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/59.jpg)
𝑓(𝑥)~𝒢𝒫(𝜇 𝑥 , 𝑘 𝑥, 𝑥′ )
Function Gaussian Process
Mean function
Covariance kernel function
𝜇 𝑥 = 𝔼 𝑓 𝑥
𝑘 𝑥, 𝑥′ = Cov 𝑓 𝑥 , 𝑓 𝑥′
[𝑓(𝑥1), … , 𝑓(𝑥𝑁)]~𝒩(𝝁, 𝚺)
Function evaluations Multivariate Gaussian
follows
Mean vector
Covariance matrix
𝝁 = [𝜇 𝑥1 , … , 𝜇 𝑥𝑁 ]
𝚺𝑖𝑗 = 𝑘 𝑥𝑖 , 𝑥𝑗
Gaussian Processes (GP)
![Page 60: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/60.jpg)
𝑓(𝑥)~𝒢𝒫(𝜇 𝑥 , 𝑘 𝑥, 𝑥′ )
GP Examples
![Page 61: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/61.jpg)
GP Base Kernels
Base
kernel
Encoding
functionKernel function
Paramet
ers
Example kernel
function shape
Example
encoded
functions
LIN(𝑥, 𝑥’)Linear
function𝜎2(𝑥 − ℓ)(𝑥′ − ℓ) 𝜎, ℓ
SE(𝑥, 𝑥’)Smooth
function𝜎2 exp −
𝑥 − 𝑥′ 2
2ℓ2𝜎, ℓ
PER(𝑥, 𝑥’)Periodic
functionIn appendix 𝜎, ℓ, 𝑝
![Page 62: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/62.jpg)
𝑓(𝑥)~𝒢𝒫(𝜇 𝑥 , 𝑘 𝑥, 𝑥′ )
Find appropriate kernel(1) Encode characteristic
Multi-kernel Learning
![Page 63: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/63.jpg)
𝑓(𝑥)~𝒢𝒫(𝜇 𝑥 , 𝑘 𝑥, 𝑥′ )
Find appropriate kernel(1) Encode characteristic
If 𝑔(𝑥)~𝒢𝒫 0, 𝑘𝑔 , ℎ 𝑥 ~𝒢𝒫 0, 𝑘ℎ and 𝑔 𝑥 ⊥ ℎ 𝑥
, then
𝑔 𝑥 + ℎ 𝑥 ~ 𝒢𝒫 0, 𝑘𝑔 + 𝑘ℎ
𝑔 𝑥 × ℎ 𝑥 ~ 𝒢𝒫 0, 𝑘𝑔 × 𝑘ℎ
(2) Compose new kernel (appendix)
The Automatic Statistician* Automatic Bayesian Covariance Discovery (http://www.automaticstatistician.com/)
Ghahramani, 2015
![Page 64: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/64.jpg)
The Automatic Statistician: Kernel Composition
Op. Concept Params ExampleExample kernel
function shape
Example encoded
functions
+
Addition
Superposition
OR operator
N/A
SE + PER
LIN + PER
×Multiplication
AND operatorN/A SE × PER
Grosse et. al., 2012
![Page 65: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/65.jpg)
Kernel Composition & Covariance Decomposition
+ = →
→ +≈
Kernel Composition: Generate Data from Models
Covariance Decomposition: Learn Explainable Models from Data
![Page 66: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/66.jpg)
Kernel Composition & Covariance Decomposition
Kernel Composition: Generate Data from Models
+ = →
Covariance Decomposition: Learn Explainable Models from Data
→ +≈
![Page 67: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/67.jpg)
(2) Optimize
(1) Expand
(3) Select
❖Iteratively select best model (structure 𝑘, parameter )
(1) Expand: the current kernel
(2) Optimize: conjugate gradient descent
(3) Select: the best kernel in the level (greedy)
(4) Iterate: get back to (1) for the next level
(2) Learning algorithm (Composite Kernel Learning)
Negative log-likelihood
Model complexity
Num. of model parameters
Num. of data points
(1) Optimization criteria: Bayesian Information Criterion (BIC)
BIC ℳ = −2 log𝑃 𝐷 ℳ + ℳ log 𝐷
The Automatic Statistician: Greedy Kernel Search
Duvenaud et. al., 2014
![Page 68: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/68.jpg)
Linear function
decrease x/week
Smooth function
Length scale: y weeks
Rapidly varying
smooth function
Length scale: z hours
The Automatic Statistician: A Sample Report
Lloyd et. al., 2014
![Page 69: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/69.jpg)
The Automatic Statistician: A Sample Report=
+ +
≈
Lloyd et. al., 2014
![Page 70: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/70.jpg)
The Automatic Statistician: Extrapolation Performance
13 regression datasets
Lloyd et. al., 2014
![Page 71: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/71.jpg)
Challenge: The Automatic StatisticianIncorporating Global Changes
Linear function
decrease x/week
Smooth function
Length scale: y weeks
Rapidly varying
smooth function
Length scale: z hours
Adjusted Close of
General Electronics
9/11, 2001
![Page 72: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/72.jpg)
Challenge: The Automatic StatisticianIncorporating Global Changes
Linear function
decrease x/week
Smooth function
Length scale: y weeks
Rapidly varying
smooth function
Length scale: z hours
Adjusted Close of
General Electronics
9/11, 2001
![Page 73: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/73.jpg)
Challenge: The Automatic StatisticianQ: How about handling multiple time series?
Adjusted Close of
General Electronics, Microsoft, ExxonMobil
9/11, 2001
• Exploit multiple time series
• Find global descriptions
• Hope better predictive
performance
![Page 74: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/74.jpg)
Constant function
Sudden drop btw
9/12/01 ~ 9/15/01
Smooth function
Length scale: y weeks
Rapidly varying
smooth function
Length scale: z hours
Problem (Our research)
Descriptive prediction of multiple time series
![Page 75: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/75.jpg)
Model: Composite Kernel Learning (The Automatic Statistician)
Fixed Grammar Optimize
Latent function evaluation
Observation
Covariance kernel functionMean function
Gaussian Process
Latent function
𝑃 𝐷 ℳ = 𝑃 𝐷 𝒢𝒫 0, 𝑘 𝑥, 𝑥′; 𝜃
A Generalized Multi Kernel Learning
![Page 76: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/76.jpg)
Model: Semi-Relational Kernel Learning
Scale parameter
Distinctive kernel
Latent function
Fixed Grammar Optimize
Latent function evaluation
Observation
Covariance kernel functionMean function
Gaussian Process
Fixed (Spectral Mixture) Optimize
𝑃 𝐷 ℳ =ෑ
𝑗=1
𝑀
𝑃 𝑑𝑗 𝒢𝒫 0, 𝜎𝑗 × 𝑘 𝑥, 𝑥′; 𝜃 + 𝑘𝑗(𝑥, 𝑥′; 𝜃𝑗)
Hwang et al., 2016
![Page 77: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/77.jpg)
Experiments on Financial Data Sets
Descriptions Graphs (normalized) Details
9 adjusted close
of stock figures
GE,MSFT, XOM,
PFE, C, WMT,
INTC, BP, AIG
6 US housing price indices
New York, Los Angeles
Chicago, Pheonix,
San Diego, San Fancisco
4 emerging market
currency exchanges
Indonesian - IDR
Malaysian - MYR
South African - ZAR
Russian - RUB
(2003 ~ 2013)
(2001 ~ 2002)
(2016)
Hwang et al., 2016
![Page 78: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/78.jpg)
Qualitative Results
Adjusted Closes Component 1 Adjusted Closes Component 2
US stock market values suddenly drop after US 9/11 attacks.
Currency exchange is affected by FED’s policy change in interest rates around middle
Sep 2015.
Automatic Statistician Relational Automatic Statistician
4 currency exchange rates Learned component
Hwang et al., 2016
![Page 79: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/79.jpg)
9 stocks 6 house price indices 4 currency exchanges
Quantitative Results
Hwang et al., 2016
![Page 80: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/80.jpg)
STOCK3 = {GE, MSFT, XOM}
STOCK6= STOCK3 + {PFE, C, WMT}
STOCK9 = STOCK6 + {INTC, BP, AIG}
HOUSE2 = {NY, LA}
HOUSE4 = HOUSE2 + {Chicago, Pheonix}
HOUSE6 = HOUSE4 + {San Diego, San Francisco}
CURRENCY4 = {IDR, MYR,ZAR,RUB}
Quantitative Results
Hwang et al., 2016
![Page 81: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/81.jpg)
An Automatically Generated Report
![Page 82: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/82.jpg)
An Automatically Generated Report
![Page 83: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/83.jpg)
An Automatically Generated Report
![Page 84: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/84.jpg)
An Automatically Generated Report
![Page 85: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/85.jpg)
An Automatically Generated Report
![Page 86: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/86.jpg)
An Automatically Generated Report
![Page 87: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/87.jpg)
An Automatically Generated Report
![Page 88: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/88.jpg)
Challenges: Selective Kernel SearchQ: Can we selectively search over time series?
![Page 89: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/89.jpg)
Discovering Explainable Latent Covariance Structures for
Multiple Time Series
Indian Buffet Processes (IBP) + Gaussian Processes(Nonparametric Clustering) (Nonparametric Regression)
Tong, Choi, 2018
![Page 90: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/90.jpg)
Discovering Explainable Latent Covariance Structures for
Multiple Time Series –Version 1
South Africa
Indonesia
Malaysia
Russia
South African Rand and Indonesian Rupiah and Malaysian Ringgit and Russian Rouble share
the following properties
→ This component is a smooth function with a typical lengthscale of 6.4 days. This
component applies until Sep. 15th 2015 and from Sep. 17th 2015 onwards.
Indonesian Rupiah and Malaysian Ringgit and Russian Rouble share the following properties
→ This component is linearly increasing.
Tong, Choi, 2018
![Page 91: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/91.jpg)
Discovering Explainable Latent Covariance Structures for
Multiple Time Series –Version II
Different realization
Tong et al., 2018
![Page 92: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/92.jpg)
Discovering Explainable Latent Covariance Structures for
Multiple Time Series Tong et al., 2018
![Page 93: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/93.jpg)
Future: Toward Reading/Explaining ReportsBeyond Chart-based Analysis
![Page 94: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/94.jpg)
Future: Toward Reading/Explaining ReportsBeyond Chart-based Analysis
![Page 95: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/95.jpg)
Future: Finding Explanation from ReportsRead the Report and Explain It
Multiple Stocks
Annual Reports
Bayesian Learning
Deep Learning
PredictionFind
Explanation
Report
![Page 96: Interpreting and Explaining Deep Neural Networks: A ... · Slides courtesy of Eamonn Keogh The dataset is an hour of EOG (eye movement) data of a sleeping patient, sampled at 100](https://reader035.fdocuments.us/reader035/viewer/2022071514/61353406dfd10f4dd73c390f/html5/thumbnails/96.jpg)
Conclusions
- Automated data collection and processing soon will change
our daily life.
- Automated narrative generation methods/frameworks may ha
ve widespread applications such as finance and media.
- Compositions of explainable models would generate more hu
man understandable descriptions of data.
- Reading and Explaining Articles (e.g., Annual Report) would gr
eatly help to improve the prediction accuracy in the future.