SlideShare a Scribd company logo
Looking beyond LSTMs: Alternatives in Time
Series Modelling using Neural Nets
Prepared exclusively for talk at Anthill.
Aditya Patel
aditya@stasislabs.com
Outline
1 Motivation
2 Hail RNN
3 Our Methodology
4 Other Methodologies
5 Takeaways
1.1 Issues with Real world data
Motivation1
Missing Data Noisy Data
*Example from healthcare data
*Expectation RealityV/s
1.2 Current State of Time Series Modelling
Motivation1
Similarity Measure Approaches
Dynamic Time Warping + KNN Classifier
Arima Modelling
ŷt = μ + ϕ1 yt-1 +…+ ϕp yt-p - θ1et-1 -…- θqet-q
Key Issues:
○ Time Complexity is high. Hard to implement in real time.
Neural Networks, especially Recurrent
Neural Nets
○ Considers linear functions in past data.
○ Doesn’t converge when applied on Noisy data or incomplete data.
• Applications:
• Music Generation
• Machine Translation
• Speech Recognition
• Application of auto encoders
2.1
Recurrent Neural Nets are all the Rage
in Sequences!!
Hail RNN2
• Recurrent Neural Nets (RNN) exploit the sequential nature of time series.
• Each output is function of a previous outputs and learns patterns via parameter sharing.
• Able to glean context by parameter sharing and unbounded nature of RNNs
2.2 Recurrent Neural Nets Highlights
Fig. 2.1 : Example of Typical RNN
• Susceptible to failure with noisy and missing data.
Not able to differentiate the two classes.(Fig. 2.3)
• Time Complexity: LSTM’s takes a long time to learn
the complex dependencies.
• Data requirements: LSTM’s require huge amount of
data to learn a reasonable representation.
2.3 Challenges with RNN
Hail RNN2
Fig. 2.3 : Example of Autoencoder from Stasis Vitals Dataset
• Highlights:
○ Takes input sequences to a kernel to create
features which calculates output taking
neighboring values into account.
○ Creates features rich in temporal history.
3.1.1 1D Convolution Neural Nets
Our Methodology - Convolutional Networks3
Equation: 1-D Convolution
* Aditya Patel et al. - http://www.scitepress.org/PublicationsDetail.aspx?ID=i4mSbjSAMY8=&t=1
*
• Dataset Used : MIMIC III [Opensource MIT].
• Training Examples : ~ 4000 patients with ~170000 Samples.
• Prediction Label : Mortality of the patient
• Dataset description : Physiological vitals ex. BP, Temp,
SpO2.
3.1.2 Training Summary*:
Our Methodology - Convolutional Networks3
* Aditya Patel et al. - http://www.scitepress.org/PublicationsDetail.aspx?ID=i4mSbjSAMY8=&t=1
*
Methods Layers #Hidden Units
LSTM 2 LSTM 128
1D CNN 4 convolution 64
Results of 1D-CNN3
• Parallel processing.
• Can glean out patterns locally.
• Works well on noisy data.
• Robust against missing data.
3.1.3 Advantages of 1D CNN
* http://www.scitepress.org/PublicationsDetail.aspx?ID=i4mSbjSAMY8=&t=1 ** On a tesla K80 GPU
1
*
Model AUC Precision Recall Time
(per epoch)
CNN 0.87 0.74 0.82 84(sec.)
RNN 0.80 0.69 0.80 494(sec.)
*
• Google - Wavenet : Generative Model for Raw Audio
• Facebook : Neural Machine Translation.
3.1.4 Other Companies employing CNN
**
Our Methodology - Convolutional Networks3
• The convolution in the architecture are causal.
• To accommodate longer history dilation was performed in layers.
*https://arxiv.org/pdf/1803.01271.pdf
*
3.2.1 Temporal Convolutional Networks (TCN)
Results - Convolutional Networks3
• Every residual block has two identical dilated causal convolutions,
𝑇(𝑛) = 𝑇(𝑛 − 1) + 2 ∗ [𝑘𝑒𝑟𝑛𝑒𝑙_𝑠𝑖𝑧𝑒(𝑛) − 1] ∗ 𝑑𝑖𝑙𝑎𝑡𝑖𝑜𝑛(𝑛)
• In our scenario, we capture 6 vital signs every 5 minutes interval. Since we are considering 2 hours
of data, it gives us 25 timestamps. 25*6 data points flattening to get a shape of (150,1).
• Best Hyperparameters : Dilation = [1,2,4] , kernel size= 3
3.2.2 Calculate Receptive Field of TCN
Table 3: Aggregated results on multi-class for predicting patient deterioration
RNN 1D CNN TCN
precision 0.77 0.83 0.81
recall 0.77 0.83 0.80
f1-score 0.74 0.83 0.80
Class of Interest Precision
TCN 0.91
1D CNN 0.83
3.3 Stasis Solution
Application3
• Quantize the data into bins.
• Convert each bins into character.
• Use regular Natural Language Modelling techniques on it.
4.2 Bag of Words Approach
Other Noteworthy Methodologies4
• Other interesting approaches:
• Transform Time Series into Bag of Words.
• Transform Time Series to Image.
4.1 Exploration
Model Precision Recall f1- score
PAA- XGB 0.74 0.74 0.74
• Transform time series to polar coordinates.
• Create Gramian matrixes :
• Gramian Angular Summation Field (GASF).
• Gramian Angular Difference Fields (GADF).
• Create image from Markov Transition Fields (MTF).
• Create image by calculating Recurrence Plots (RP)
• Feed the images to tiled CNN’s.
4.3 Transform Time Series to Images
Time Series
MTFGASF
GADF RP
Model Precision Recall f1- score
GADF + CNN 0.80 0.80 0.80
Other Noteworthy Methodologies4
Takeaways5
5.1
• Understanding the problem is as critical as using the latest algorithms.
• With the advent of new techniques in RNN’s (example: Attention mechanisms) some of
the drawbacks mentioned in the presentation have been rectified, but we wanted to
showcase that it’s not a cure-all bullet as perceived to be.
• We need to explore existing algorithms intelligently to create robust solutions.
• 1D CNN performs well in realistic settings.
• More research is required to improve upon restrictions of Convolutional Networks for
Sequence Prediction Tasks.
5.1
Q & A
Aditya Patel | Head of Data Science
aditya@stasislabs.com | +91 9108932213
© Copyright. 2018. Stasis Labs Inc./ Stasis Health Private Limited. All Rights Reserved.
Appendix
3Wikipedia
3*
Standard EWS Evaluation

More Related Content

What's hot

DNR - Auto deep lab paper review ppt
DNR - Auto deep lab paper review pptDNR - Auto deep lab paper review ppt
DNR - Auto deep lab paper review ppt
taeseon ryu
 
Learning multifractal structure in large networks (Purdue ML Seminar)
Learning multifractal structure in large networks (Purdue ML Seminar)Learning multifractal structure in large networks (Purdue ML Seminar)
Learning multifractal structure in large networks (Purdue ML Seminar)
Austin Benson
 
Object Detection Using R-CNN Deep Learning Framework
Object Detection Using R-CNN Deep Learning FrameworkObject Detection Using R-CNN Deep Learning Framework
Object Detection Using R-CNN Deep Learning Framework
Nader Karimi
 
Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)
Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)
Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)
Universitat Politècnica de Catalunya
 
Trackster Pruning at the CMS High-Granularity Calorimeter
Trackster Pruning at the CMS High-Granularity CalorimeterTrackster Pruning at the CMS High-Granularity Calorimeter
Trackster Pruning at the CMS High-Granularity Calorimeter
Yousef Fadila
 
Detection focal loss 딥러닝 논문읽기 모임 발표자료
Detection focal loss 딥러닝 논문읽기 모임 발표자료Detection focal loss 딥러닝 논문읽기 모임 발표자료
Detection focal loss 딥러닝 논문읽기 모임 발표자료
taeseon ryu
 
K means clustering algorithm
K means clustering algorithmK means clustering algorithm
K means clustering algorithm
Darshak Mehta
 
Focal loss for dense object detection
Focal loss for dense object detectionFocal loss for dense object detection
Focal loss for dense object detection
DaeHeeKim31
 
Support Vector Machine (Classification) - Step by Step
Support Vector Machine (Classification) - Step by StepSupport Vector Machine (Classification) - Step by Step
Support Vector Machine (Classification) - Step by Step
Manish nath choudhary
 
virtualization
virtualizationvirtualization
virtualization
Avi Nash
 
LIDAR- Light Detection and Ranging.
LIDAR- Light Detection and Ranging.LIDAR- Light Detection and Ranging.
LIDAR- Light Detection and Ranging.
Gaurav Agarwal
 
Advanced deep learning based object detection methods
Advanced deep learning based object detection methodsAdvanced deep learning based object detection methods
Advanced deep learning based object detection methods
Brodmann17
 
05 k-means clustering
05 k-means clustering05 k-means clustering
05 k-means clustering
Subhas Kumar Ghosh
 
Autoencoder
AutoencoderAutoencoder
Autoencoder
Mehrnaz Faraz
 
computer networking
computer networkingcomputer networking
computer networking
Avi Nash
 
(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...
(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...
(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...
e8xu
 
Data structue q & a
Data structue q & aData structue q & a
Data structue q & a
rohanbajaj222
 
Single Shot Multibox Detector
Single Shot Multibox DetectorSingle Shot Multibox Detector
Single Shot Multibox Detector
NamHyuk Ahn
 
presentation
presentationpresentation
presentationjie ren
 

What's hot (20)

DNR - Auto deep lab paper review ppt
DNR - Auto deep lab paper review pptDNR - Auto deep lab paper review ppt
DNR - Auto deep lab paper review ppt
 
Learning multifractal structure in large networks (Purdue ML Seminar)
Learning multifractal structure in large networks (Purdue ML Seminar)Learning multifractal structure in large networks (Purdue ML Seminar)
Learning multifractal structure in large networks (Purdue ML Seminar)
 
Object Detection Using R-CNN Deep Learning Framework
Object Detection Using R-CNN Deep Learning FrameworkObject Detection Using R-CNN Deep Learning Framework
Object Detection Using R-CNN Deep Learning Framework
 
Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)
Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)
Object Detection (D2L5 Insight@DCU Machine Learning Workshop 2017)
 
Trackster Pruning at the CMS High-Granularity Calorimeter
Trackster Pruning at the CMS High-Granularity CalorimeterTrackster Pruning at the CMS High-Granularity Calorimeter
Trackster Pruning at the CMS High-Granularity Calorimeter
 
Detection focal loss 딥러닝 논문읽기 모임 발표자료
Detection focal loss 딥러닝 논문읽기 모임 발표자료Detection focal loss 딥러닝 논문읽기 모임 발표자료
Detection focal loss 딥러닝 논문읽기 모임 발표자료
 
K means clustering algorithm
K means clustering algorithmK means clustering algorithm
K means clustering algorithm
 
Focal loss for dense object detection
Focal loss for dense object detectionFocal loss for dense object detection
Focal loss for dense object detection
 
Support Vector Machine (Classification) - Step by Step
Support Vector Machine (Classification) - Step by StepSupport Vector Machine (Classification) - Step by Step
Support Vector Machine (Classification) - Step by Step
 
virtualization
virtualizationvirtualization
virtualization
 
LIDAR- Light Detection and Ranging.
LIDAR- Light Detection and Ranging.LIDAR- Light Detection and Ranging.
LIDAR- Light Detection and Ranging.
 
Advanced deep learning based object detection methods
Advanced deep learning based object detection methodsAdvanced deep learning based object detection methods
Advanced deep learning based object detection methods
 
05 k-means clustering
05 k-means clustering05 k-means clustering
05 k-means clustering
 
Autoencoder
AutoencoderAutoencoder
Autoencoder
 
Kmeans
KmeansKmeans
Kmeans
 
computer networking
computer networkingcomputer networking
computer networking
 
(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...
(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...
(Paper note) Real time rgb-d camera relocalization via randomized ferns for k...
 
Data structue q & a
Data structue q & aData structue q & a
Data structue q & a
 
Single Shot Multibox Detector
Single Shot Multibox DetectorSingle Shot Multibox Detector
Single Shot Multibox Detector
 
presentation
presentationpresentation
presentation
 

Similar to Anthill Talk Aditya

Sequence Model pytorch at colab with gpu.pdf
Sequence Model pytorch at colab with gpu.pdfSequence Model pytorch at colab with gpu.pdf
Sequence Model pytorch at colab with gpu.pdf
FEG
 
Real-Time Streaming Data Analysis with HTM
Real-Time Streaming Data Analysis with HTMReal-Time Streaming Data Analysis with HTM
Real-Time Streaming Data Analysis with HTM
Numenta
 
240219_RNN, LSTM code.pptxdddddddddddddddd
240219_RNN, LSTM code.pptxdddddddddddddddd240219_RNN, LSTM code.pptxdddddddddddddddd
240219_RNN, LSTM code.pptxdddddddddddddddd
ssuser2624f71
 
Classification of indoor actions through deep neural networks
Classification of indoor actions through deep neural networksClassification of indoor actions through deep neural networks
Classification of indoor actions through deep neural networks
Cognitive Robotics and Social Sensing Lab - CNR - ICAR
 
Recurrent Neural Networks, LSTM and GRU
Recurrent Neural Networks, LSTM and GRURecurrent Neural Networks, LSTM and GRU
Recurrent Neural Networks, LSTM and GRU
ananth
 
Teach a neural network to read handwriting
Teach a neural network to read handwritingTeach a neural network to read handwriting
Teach a neural network to read handwriting
Vipul Kaushal
 
17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx
17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx
17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx
MahmoudAbuGhali
 
Data streaming algorithms
Data streaming algorithmsData streaming algorithms
Data streaming algorithms
Sandeep Joshi
 
Introduction to deep learning
Introduction to deep learningIntroduction to deep learning
Introduction to deep learning
Junaid Bhat
 
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
ssuser9357dd
 
Tsinghua University: Two Exemplary Applications in China
Tsinghua University: Two Exemplary Applications in ChinaTsinghua University: Two Exemplary Applications in China
Tsinghua University: Two Exemplary Applications in China
DataStax Academy
 
Application of machine learning and cognitive computing in intrusion detectio...
Application of machine learning and cognitive computing in intrusion detectio...Application of machine learning and cognitive computing in intrusion detectio...
Application of machine learning and cognitive computing in intrusion detectio...
Mahdi Hosseini Moghaddam
 
Neural Networks in Data Mining - “An Overview”
Neural Networks  in Data Mining -   “An Overview”Neural Networks  in Data Mining -   “An Overview”
Neural Networks in Data Mining - “An Overview”
Dr.(Mrs).Gethsiyal Augasta
 
rsec2a-2016-jheaton-morning
rsec2a-2016-jheaton-morningrsec2a-2016-jheaton-morning
rsec2a-2016-jheaton-morningJeff Heaton
 
Unsupervised Learning Clustering KMean and Hirarchical.pptx
Unsupervised Learning Clustering KMean and Hirarchical.pptxUnsupervised Learning Clustering KMean and Hirarchical.pptx
Unsupervised Learning Clustering KMean and Hirarchical.pptx
FaridAliMousa1
 
The study on mining temporal patterns and related applications in dynamic soc...
The study on mining temporal patterns and related applications in dynamic soc...The study on mining temporal patterns and related applications in dynamic soc...
The study on mining temporal patterns and related applications in dynamic soc...Thanh Hieu
 
Deep Learning
Deep LearningDeep Learning
Deep Learning
Pierre de Lacaze
 
Deep Learning Tomography
Deep Learning TomographyDeep Learning Tomography
Deep Learning Tomography
Amir Adler
 
Bioinfo ngs data format visualization v2
Bioinfo ngs data format visualization v2Bioinfo ngs data format visualization v2
Bioinfo ngs data format visualization v2
Li Shen
 
Enterprise Scale Topological Data Analysis Using Spark
Enterprise Scale Topological Data Analysis Using SparkEnterprise Scale Topological Data Analysis Using Spark
Enterprise Scale Topological Data Analysis Using Spark
Alpine Data
 

Similar to Anthill Talk Aditya (20)

Sequence Model pytorch at colab with gpu.pdf
Sequence Model pytorch at colab with gpu.pdfSequence Model pytorch at colab with gpu.pdf
Sequence Model pytorch at colab with gpu.pdf
 
Real-Time Streaming Data Analysis with HTM
Real-Time Streaming Data Analysis with HTMReal-Time Streaming Data Analysis with HTM
Real-Time Streaming Data Analysis with HTM
 
240219_RNN, LSTM code.pptxdddddddddddddddd
240219_RNN, LSTM code.pptxdddddddddddddddd240219_RNN, LSTM code.pptxdddddddddddddddd
240219_RNN, LSTM code.pptxdddddddddddddddd
 
Classification of indoor actions through deep neural networks
Classification of indoor actions through deep neural networksClassification of indoor actions through deep neural networks
Classification of indoor actions through deep neural networks
 
Recurrent Neural Networks, LSTM and GRU
Recurrent Neural Networks, LSTM and GRURecurrent Neural Networks, LSTM and GRU
Recurrent Neural Networks, LSTM and GRU
 
Teach a neural network to read handwriting
Teach a neural network to read handwritingTeach a neural network to read handwriting
Teach a neural network to read handwriting
 
17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx
17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx
17_00-Dima-Panchenko-cnn-tips-and-tricks.pptx
 
Data streaming algorithms
Data streaming algorithmsData streaming algorithms
Data streaming algorithms
 
Introduction to deep learning
Introduction to deep learningIntroduction to deep learning
Introduction to deep learning
 
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
CTF: Anomaly Detection in High-Dimensional Time Series with Coarse-to-Fine Mo...
 
Tsinghua University: Two Exemplary Applications in China
Tsinghua University: Two Exemplary Applications in ChinaTsinghua University: Two Exemplary Applications in China
Tsinghua University: Two Exemplary Applications in China
 
Application of machine learning and cognitive computing in intrusion detectio...
Application of machine learning and cognitive computing in intrusion detectio...Application of machine learning and cognitive computing in intrusion detectio...
Application of machine learning and cognitive computing in intrusion detectio...
 
Neural Networks in Data Mining - “An Overview”
Neural Networks  in Data Mining -   “An Overview”Neural Networks  in Data Mining -   “An Overview”
Neural Networks in Data Mining - “An Overview”
 
rsec2a-2016-jheaton-morning
rsec2a-2016-jheaton-morningrsec2a-2016-jheaton-morning
rsec2a-2016-jheaton-morning
 
Unsupervised Learning Clustering KMean and Hirarchical.pptx
Unsupervised Learning Clustering KMean and Hirarchical.pptxUnsupervised Learning Clustering KMean and Hirarchical.pptx
Unsupervised Learning Clustering KMean and Hirarchical.pptx
 
The study on mining temporal patterns and related applications in dynamic soc...
The study on mining temporal patterns and related applications in dynamic soc...The study on mining temporal patterns and related applications in dynamic soc...
The study on mining temporal patterns and related applications in dynamic soc...
 
Deep Learning
Deep LearningDeep Learning
Deep Learning
 
Deep Learning Tomography
Deep Learning TomographyDeep Learning Tomography
Deep Learning Tomography
 
Bioinfo ngs data format visualization v2
Bioinfo ngs data format visualization v2Bioinfo ngs data format visualization v2
Bioinfo ngs data format visualization v2
 
Enterprise Scale Topological Data Analysis Using Spark
Enterprise Scale Topological Data Analysis Using SparkEnterprise Scale Topological Data Analysis Using Spark
Enterprise Scale Topological Data Analysis Using Spark
 

Recently uploaded

Influence of Marketing Strategy and Market Competition on Business Plan
Influence of Marketing Strategy and Market Competition on Business PlanInfluence of Marketing Strategy and Market Competition on Business Plan
Influence of Marketing Strategy and Market Competition on Business Plan
jerlynmaetalle
 
一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理
v3tuleee
 
The Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series DatabaseThe Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series Database
javier ramirez
 
【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】
【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】
【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】
NABLAS株式会社
 
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
oz8q3jxlp
 
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdfCh03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
haila53
 
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptxData_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
AnirbanRoy608946
 
做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样
做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样
做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样
axoqas
 
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
ahzuo
 
Adjusting OpenMP PageRank : SHORT REPORT / NOTES
Adjusting OpenMP PageRank : SHORT REPORT / NOTESAdjusting OpenMP PageRank : SHORT REPORT / NOTES
Adjusting OpenMP PageRank : SHORT REPORT / NOTES
Subhajit Sahu
 
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
g4dpvqap0
 
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
John Andrews
 
一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理
一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理
一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理
dwreak4tg
 
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
ahzuo
 
Machine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptxMachine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptx
balafet
 
一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单
ewymefz
 
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Subhajit Sahu
 
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
Timothy Spann
 
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
74nqk8xf
 
Q1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year ReboundQ1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year Rebound
Oppotus
 

Recently uploaded (20)

Influence of Marketing Strategy and Market Competition on Business Plan
Influence of Marketing Strategy and Market Competition on Business PlanInfluence of Marketing Strategy and Market Competition on Business Plan
Influence of Marketing Strategy and Market Competition on Business Plan
 
一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理
 
The Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series DatabaseThe Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series Database
 
【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】
【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】
【社内勉強会資料_Octo: An Open-Source Generalist Robot Policy】
 
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
 
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdfCh03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
 
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptxData_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
 
做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样
做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样
做(mqu毕业证书)麦考瑞大学毕业证硕士文凭证书学费发票原版一模一样
 
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
 
Adjusting OpenMP PageRank : SHORT REPORT / NOTES
Adjusting OpenMP PageRank : SHORT REPORT / NOTESAdjusting OpenMP PageRank : SHORT REPORT / NOTES
Adjusting OpenMP PageRank : SHORT REPORT / NOTES
 
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
 
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
 
一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理
一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理
一比一原版(BCU毕业证书)伯明翰城市大学毕业证如何办理
 
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
 
Machine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptxMachine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptx
 
一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单
 
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
 
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
 
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
 
Q1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year ReboundQ1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year Rebound
 

Anthill Talk Aditya

  • 1. Looking beyond LSTMs: Alternatives in Time Series Modelling using Neural Nets Prepared exclusively for talk at Anthill. Aditya Patel aditya@stasislabs.com
  • 2. Outline 1 Motivation 2 Hail RNN 3 Our Methodology 4 Other Methodologies 5 Takeaways
  • 3. 1.1 Issues with Real world data Motivation1 Missing Data Noisy Data *Example from healthcare data *Expectation RealityV/s
  • 4. 1.2 Current State of Time Series Modelling Motivation1 Similarity Measure Approaches Dynamic Time Warping + KNN Classifier Arima Modelling ŷt = μ + ϕ1 yt-1 +…+ ϕp yt-p - θ1et-1 -…- θqet-q Key Issues: ○ Time Complexity is high. Hard to implement in real time. Neural Networks, especially Recurrent Neural Nets ○ Considers linear functions in past data. ○ Doesn’t converge when applied on Noisy data or incomplete data.
  • 5. • Applications: • Music Generation • Machine Translation • Speech Recognition • Application of auto encoders 2.1 Recurrent Neural Nets are all the Rage in Sequences!! Hail RNN2 • Recurrent Neural Nets (RNN) exploit the sequential nature of time series. • Each output is function of a previous outputs and learns patterns via parameter sharing. • Able to glean context by parameter sharing and unbounded nature of RNNs 2.2 Recurrent Neural Nets Highlights Fig. 2.1 : Example of Typical RNN
  • 6. • Susceptible to failure with noisy and missing data. Not able to differentiate the two classes.(Fig. 2.3) • Time Complexity: LSTM’s takes a long time to learn the complex dependencies. • Data requirements: LSTM’s require huge amount of data to learn a reasonable representation. 2.3 Challenges with RNN Hail RNN2 Fig. 2.3 : Example of Autoencoder from Stasis Vitals Dataset
  • 7. • Highlights: ○ Takes input sequences to a kernel to create features which calculates output taking neighboring values into account. ○ Creates features rich in temporal history. 3.1.1 1D Convolution Neural Nets Our Methodology - Convolutional Networks3 Equation: 1-D Convolution * Aditya Patel et al. - http://www.scitepress.org/PublicationsDetail.aspx?ID=i4mSbjSAMY8=&t=1 *
  • 8. • Dataset Used : MIMIC III [Opensource MIT]. • Training Examples : ~ 4000 patients with ~170000 Samples. • Prediction Label : Mortality of the patient • Dataset description : Physiological vitals ex. BP, Temp, SpO2. 3.1.2 Training Summary*: Our Methodology - Convolutional Networks3 * Aditya Patel et al. - http://www.scitepress.org/PublicationsDetail.aspx?ID=i4mSbjSAMY8=&t=1 * Methods Layers #Hidden Units LSTM 2 LSTM 128 1D CNN 4 convolution 64
  • 9. Results of 1D-CNN3 • Parallel processing. • Can glean out patterns locally. • Works well on noisy data. • Robust against missing data. 3.1.3 Advantages of 1D CNN * http://www.scitepress.org/PublicationsDetail.aspx?ID=i4mSbjSAMY8=&t=1 ** On a tesla K80 GPU 1 * Model AUC Precision Recall Time (per epoch) CNN 0.87 0.74 0.82 84(sec.) RNN 0.80 0.69 0.80 494(sec.) * • Google - Wavenet : Generative Model for Raw Audio • Facebook : Neural Machine Translation. 3.1.4 Other Companies employing CNN **
  • 10. Our Methodology - Convolutional Networks3 • The convolution in the architecture are causal. • To accommodate longer history dilation was performed in layers. *https://arxiv.org/pdf/1803.01271.pdf * 3.2.1 Temporal Convolutional Networks (TCN)
  • 11. Results - Convolutional Networks3 • Every residual block has two identical dilated causal convolutions, 𝑇(𝑛) = 𝑇(𝑛 − 1) + 2 ∗ [𝑘𝑒𝑟𝑛𝑒𝑙_𝑠𝑖𝑧𝑒(𝑛) − 1] ∗ 𝑑𝑖𝑙𝑎𝑡𝑖𝑜𝑛(𝑛) • In our scenario, we capture 6 vital signs every 5 minutes interval. Since we are considering 2 hours of data, it gives us 25 timestamps. 25*6 data points flattening to get a shape of (150,1). • Best Hyperparameters : Dilation = [1,2,4] , kernel size= 3 3.2.2 Calculate Receptive Field of TCN Table 3: Aggregated results on multi-class for predicting patient deterioration RNN 1D CNN TCN precision 0.77 0.83 0.81 recall 0.77 0.83 0.80 f1-score 0.74 0.83 0.80 Class of Interest Precision TCN 0.91 1D CNN 0.83
  • 13. • Quantize the data into bins. • Convert each bins into character. • Use regular Natural Language Modelling techniques on it. 4.2 Bag of Words Approach Other Noteworthy Methodologies4 • Other interesting approaches: • Transform Time Series into Bag of Words. • Transform Time Series to Image. 4.1 Exploration Model Precision Recall f1- score PAA- XGB 0.74 0.74 0.74
  • 14. • Transform time series to polar coordinates. • Create Gramian matrixes : • Gramian Angular Summation Field (GASF). • Gramian Angular Difference Fields (GADF). • Create image from Markov Transition Fields (MTF). • Create image by calculating Recurrence Plots (RP) • Feed the images to tiled CNN’s. 4.3 Transform Time Series to Images Time Series MTFGASF GADF RP Model Precision Recall f1- score GADF + CNN 0.80 0.80 0.80 Other Noteworthy Methodologies4
  • 15. Takeaways5 5.1 • Understanding the problem is as critical as using the latest algorithms. • With the advent of new techniques in RNN’s (example: Attention mechanisms) some of the drawbacks mentioned in the presentation have been rectified, but we wanted to showcase that it’s not a cure-all bullet as perceived to be. • We need to explore existing algorithms intelligently to create robust solutions. • 1D CNN performs well in realistic settings. • More research is required to improve upon restrictions of Convolutional Networks for Sequence Prediction Tasks.
  • 17. Aditya Patel | Head of Data Science aditya@stasislabs.com | +91 9108932213 © Copyright. 2018. Stasis Labs Inc./ Stasis Health Private Limited. All Rights Reserved.