SlideShare a Scribd company logo
1 of 12
Neural Network Hyperparameter
Optimization using Q-Learning
Oleksii Volkovskyi
American Heritage School
Neural Network Background
● A statistical model
○ Given Inputs, computes an Output
○ Inspired by the Human Brain
○ Consists computational Units with Parameters
■ Uses complex nonlinear functions in order to
achieve high results
○ Relies on computational power and data size
○ Backpropagation can be used to “train” the network
■ Works on the principle of comparing current
output to the dataset output and self
modification in order to increase accuracy
Neural Network Hyperparameters
● Hyperparameters
○ Constants chosen before Training
● Regularisation Constant (Alpha)
○ Throttles complexity of trained model
○ Small constant values results in simpler models
○ Is generally randomly selected
● Size of Hidden Layers
○ Affects computational complexity, and can have a
negative effect
○ Is generally a human decision
○ Computational Expense
● Random selection and arbitrary decision can lead to
suboptimal performance and failure to reach the global
minimum of the neural network (bottom right graph)
Q Learning Algorithm
● In order to tackle the problem of suboptimal
hyperparameters, a reinforcement learning algorithm can
be used
○ Q learning is an algorithm that makes decision given a
state
○ The algorithm defines a Q-matrix that acts as a
decision matrix
■ The algorithm iteratively improves the Q-matrix
given a training dataset
○ Q learning can be used with a dynamic Q-matrix that
can be modified when a new state is created, which
means that the number of states doesn’t have a limit
Q Learning Algorithm
● Q - Learning will be applied to the Neural Network Hyperparameter Selection task as
following:
○ The Q-Matrix will be a multidimensional array, depending on the number of
hyperparameters and chosen states
■ In this case, the chosen hyperparameter was the regularisation constant
■ The state will be defined as an array of the following:
● The current regularisation constant value
● The current bias-variance metrics
○ The reward will be computed as the F-score of the algorithm (accuracy metric)
○ When trained on multiple networks, the algorithm should be able to generalise to any
Neural Network
■ The goal will be to converge to Optimal Hyperparameters in shorter computational
time
Hypothesis
● Q-learning, with hyperparameters and bias-variance metrics given as states and the
F-score as reward, can be used to develop a general algorithm for tuning the hyper-
parameters of a neural network.
● General neural network algorithms will:
○ Eliminate the need for human interference in neural network algorithm
training
○ Maximise the capabilities of the neural network concept
Bias Variance Metrics and F-Score
● Bias and Variance Metrics will be defined by the given formulas
○ They are responsible for the intuition behind Q-learning
algorithm
○ Bias - The ability of an algorithm to fit training data
○ Variance - The ability of an algorithm to generalise to
test data
○ Given these metrics, the Q-learning system is tasked with
scaling the regularisation constant
■ Increasing - Simpler system
■ Decreasing - Complex System
● F-score is a more rigorous alternative to accuracy
○ Handles skewed datasets more strictly
○ Eliminates the need for large pre-processing of data
Procedure
1. Create a Neural Network Function, that Inputs Regularisation Alpha, Outputs F-score, Bias, and
Variance.
2. Write a function that selects action based on Q-matrix Reward and euclidean distance to possible
hyperparameter values
3. Write the Q-learning iterative learning loop, that updates the Q-matrix based on the computed
reward and action chosen
4. Train Algorithm on Train Data, which I have chosen to be a poisonous mushroom classification
dataset
5. Benchmark algorithm on Test Data
a. Record Time and Final Accuracy of Algorithm (Average across 10 Runs)
6. Run Random Selection algorithm on Test Data as Control (Average data from 10 Runs)
a. Since random selection can narrow down to an infinitely small range of hyperparameters, the
results will compare computational time, while accuracy will be a controlled variable
Results
● Q-learning yielded a decrease in Average Computational Time
when compared to Random Selection
○ 28.65% on the test dataset
○ 47.05% on the train dataset
● It was given the task to generalise from training to test datasets
○ The datasets had similar concepts, yet entirely different
tasks
■ Safe/Poisonous Mushroom Classification (Train)
■ Malignant/Benign Cancer Classification (Test)
○ The algorithm was prone to overfitting on the train dataset
■ The algorithm was trained on one dataset due to
limitations in hardware and time
Conclusion
● The hypothesis was validated
○ The Q-Learning algorithm was able to fit
hyperparameters when given an unknown dataset
■ Additionally, the Q-Learning algorithm
Outperformed baseline algorithm (Random
Selection) in terms of computational time
○ The improvements in computational time were made
due to proper selection of hyperparameters, without
modifying the core of the neural network
Future Improvements
○ Large improvements can be gained from more complex reinforcement
learning algorithms
■ One example is Deep Q-Learning, used by Google in the development
of AlphaGo
○ Higher variety of Hyperparameters would validate the experiment further
■ Hyperparameters such as hidden layer size and amount of layers could
be used, as well as the choice of computational unit function
■ This would come with an increase in training time and overall
computational time, because the algorithm would have to consider
exponentially more options
○ Training on a variety of datasets would eliminate overfitting
Applications
● General neural network algorithms would:
○ Eliminate the need for human interference
○ Maximise Performance of neural networks
○ Have a chance to reach Bayes’ Optimal Error, which is defined as the smallest error
possible for a function to have on a dataset
● Hyperparameter training is applicable to other, more complex, types of Neural Networks
○ Convolutional Neural Networks
■ Autonomous Driving
■ Face Recognition / Verification
○ Recurrent Neural Networks
■ Speech Recognition
■ Music Composition

More Related Content

What's hot

Fitness Inheritance in Evolutionary and
Fitness Inheritance in Evolutionary andFitness Inheritance in Evolutionary and
Fitness Inheritance in Evolutionary andPier Luca Lanzi
 
Real-time PMU Data Recovery Application Based on Singular Value Decomposition
Real-time PMU Data Recovery Application Based on Singular Value DecompositionReal-time PMU Data Recovery Application Based on Singular Value Decomposition
Real-time PMU Data Recovery Application Based on Singular Value DecompositionPower System Operation
 
Kcc201728apr2017 170828235330
Kcc201728apr2017 170828235330Kcc201728apr2017 170828235330
Kcc201728apr2017 170828235330JEE HYUN PARK
 
Electricity Price Forecasting Using ELM-Tree Approach
Electricity Price Forecasting Using ELM-Tree ApproachElectricity Price Forecasting Using ELM-Tree Approach
Electricity Price Forecasting Using ELM-Tree ApproachIRJET Journal
 
Neural networks for the prediction and forecasting of water resources variables
Neural networks for the prediction and forecasting of water resources variablesNeural networks for the prediction and forecasting of water resources variables
Neural networks for the prediction and forecasting of water resources variablesJonathan D'Cruz
 
Neural Network Presentation
Neural Network PresentationNeural Network Presentation
Neural Network PresentationOmoye
 
A methodology for full system power modeling in heterogeneous data centers
A methodology for full system power modeling in  heterogeneous data centersA methodology for full system power modeling in  heterogeneous data centers
A methodology for full system power modeling in heterogeneous data centersRaimon Bosch
 
Trajectory Control With MPC For A Robot Manipülatör Using ANN Model
Trajectory Control With MPC For A Robot Manipülatör Using  ANN ModelTrajectory Control With MPC For A Robot Manipülatör Using  ANN Model
Trajectory Control With MPC For A Robot Manipülatör Using ANN ModelIJMER
 

What's hot (15)

Chap2 slides
Chap2 slidesChap2 slides
Chap2 slides
 
Fitness Inheritance in Evolutionary and
Fitness Inheritance in Evolutionary andFitness Inheritance in Evolutionary and
Fitness Inheritance in Evolutionary and
 
Real-time PMU Data Recovery Application Based on Singular Value Decomposition
Real-time PMU Data Recovery Application Based on Singular Value DecompositionReal-time PMU Data Recovery Application Based on Singular Value Decomposition
Real-time PMU Data Recovery Application Based on Singular Value Decomposition
 
Chap1 slides
Chap1 slidesChap1 slides
Chap1 slides
 
Kcc201728apr2017 170828235330
Kcc201728apr2017 170828235330Kcc201728apr2017 170828235330
Kcc201728apr2017 170828235330
 
nnUNet
nnUNetnnUNet
nnUNet
 
Electricity Price Forecasting Using ELM-Tree Approach
Electricity Price Forecasting Using ELM-Tree ApproachElectricity Price Forecasting Using ELM-Tree Approach
Electricity Price Forecasting Using ELM-Tree Approach
 
Neural networks for the prediction and forecasting of water resources variables
Neural networks for the prediction and forecasting of water resources variablesNeural networks for the prediction and forecasting of water resources variables
Neural networks for the prediction and forecasting of water resources variables
 
Neural Network Presentation
Neural Network PresentationNeural Network Presentation
Neural Network Presentation
 
A methodology for full system power modeling in heterogeneous data centers
A methodology for full system power modeling in  heterogeneous data centersA methodology for full system power modeling in  heterogeneous data centers
A methodology for full system power modeling in heterogeneous data centers
 
Trajectory Control With MPC For A Robot Manipülatör Using ANN Model
Trajectory Control With MPC For A Robot Manipülatör Using  ANN ModelTrajectory Control With MPC For A Robot Manipülatör Using  ANN Model
Trajectory Control With MPC For A Robot Manipülatör Using ANN Model
 
Mnist soln
Mnist solnMnist soln
Mnist soln
 
Chap5 slides
Chap5 slidesChap5 slides
Chap5 slides
 
Fulltext
FulltextFulltext
Fulltext
 
Grds conferences icst and icbelsh (5)
Grds conferences icst and icbelsh (5)Grds conferences icst and icbelsh (5)
Grds conferences icst and icbelsh (5)
 

Similar to Sigma Xi Research Showcase 2018 - Oleksii Volkovskyi

Artificial Intelligence Chapter 9 Negnevitsky
Artificial Intelligence Chapter 9 NegnevitskyArtificial Intelligence Chapter 9 Negnevitsky
Artificial Intelligence Chapter 9 Negnevitskylopanath
 
Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...
Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...
Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...Universitat Politècnica de Catalunya
 
Deep learning crash course
Deep learning crash courseDeep learning crash course
Deep learning crash courseVishwas N
 
Deep Dive into Hyperparameter Tuning
Deep Dive into Hyperparameter TuningDeep Dive into Hyperparameter Tuning
Deep Dive into Hyperparameter TuningShubhmay Potdar
 
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHMJOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHMmailjkb
 
Framework for Inter-Model Analysis of Cyber-Physical Systems
Framework for Inter-Model Analysis of Cyber-Physical SystemsFramework for Inter-Model Analysis of Cyber-Physical Systems
Framework for Inter-Model Analysis of Cyber-Physical SystemsIvan Ruchkin
 
Policy Based reinforcement Learning for time series Anomaly detection
Policy Based reinforcement Learning for time series Anomaly detectionPolicy Based reinforcement Learning for time series Anomaly detection
Policy Based reinforcement Learning for time series Anomaly detectionKishor Datta Gupta
 
Incremental Machine Learning.pptx
Incremental Machine Learning.pptxIncremental Machine Learning.pptx
Incremental Machine Learning.pptxSHAILIPATEL19
 
A Framework for Scene Recognition Using Convolutional Neural Network as Featu...
A Framework for Scene Recognition Using Convolutional Neural Network as Featu...A Framework for Scene Recognition Using Convolutional Neural Network as Featu...
A Framework for Scene Recognition Using Convolutional Neural Network as Featu...Tahmid Abtahi
 
Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...
Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...
Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...Databricks
 
Statistical Arbitrage
Statistical ArbitrageStatistical Arbitrage
Statistical ArbitrageShubham Patil
 
Netflix SRE perf meetup_slides
Netflix SRE perf meetup_slidesNetflix SRE perf meetup_slides
Netflix SRE perf meetup_slidesEd Hunter
 
An Introduction to Deep Learning
An Introduction to Deep LearningAn Introduction to Deep Learning
An Introduction to Deep Learningmilad abbasi
 
Introduction to Deep Learning
Introduction to Deep LearningIntroduction to Deep Learning
Introduction to Deep LearningMehrnaz Faraz
 
Using SigOpt to Tune Deep Learning Models with Nervana Cloud
Using SigOpt to Tune Deep Learning Models with Nervana CloudUsing SigOpt to Tune Deep Learning Models with Nervana Cloud
Using SigOpt to Tune Deep Learning Models with Nervana CloudSigOpt
 

Similar to Sigma Xi Research Showcase 2018 - Oleksii Volkovskyi (20)

Artificial Intelligence Chapter 9 Negnevitsky
Artificial Intelligence Chapter 9 NegnevitskyArtificial Intelligence Chapter 9 Negnevitsky
Artificial Intelligence Chapter 9 Negnevitsky
 
Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...
Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...
Training Deep Networks with Backprop (D1L4 Insight@DCU Machine Learning Works...
 
Grad presentation
Grad presentationGrad presentation
Grad presentation
 
Deep learning crash course
Deep learning crash courseDeep learning crash course
Deep learning crash course
 
Deep Dive into Hyperparameter Tuning
Deep Dive into Hyperparameter TuningDeep Dive into Hyperparameter Tuning
Deep Dive into Hyperparameter Tuning
 
iterativealgorithms.ppsx
iterativealgorithms.ppsxiterativealgorithms.ppsx
iterativealgorithms.ppsx
 
Iterative Algorithms.ppsx
Iterative Algorithms.ppsxIterative Algorithms.ppsx
Iterative Algorithms.ppsx
 
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHMJOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
 
KCC2017 28APR2017
KCC2017 28APR2017KCC2017 28APR2017
KCC2017 28APR2017
 
Framework for Inter-Model Analysis of Cyber-Physical Systems
Framework for Inter-Model Analysis of Cyber-Physical SystemsFramework for Inter-Model Analysis of Cyber-Physical Systems
Framework for Inter-Model Analysis of Cyber-Physical Systems
 
Policy Based reinforcement Learning for time series Anomaly detection
Policy Based reinforcement Learning for time series Anomaly detectionPolicy Based reinforcement Learning for time series Anomaly detection
Policy Based reinforcement Learning for time series Anomaly detection
 
Incremental Machine Learning.pptx
Incremental Machine Learning.pptxIncremental Machine Learning.pptx
Incremental Machine Learning.pptx
 
A Framework for Scene Recognition Using Convolutional Neural Network as Featu...
A Framework for Scene Recognition Using Convolutional Neural Network as Featu...A Framework for Scene Recognition Using Convolutional Neural Network as Featu...
A Framework for Scene Recognition Using Convolutional Neural Network as Featu...
 
C3 w3
C3 w3C3 w3
C3 w3
 
Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...
Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...
Apache Spark Based Hyper-Parameter Selection and Adaptive Model Tuning for De...
 
Statistical Arbitrage
Statistical ArbitrageStatistical Arbitrage
Statistical Arbitrage
 
Netflix SRE perf meetup_slides
Netflix SRE perf meetup_slidesNetflix SRE perf meetup_slides
Netflix SRE perf meetup_slides
 
An Introduction to Deep Learning
An Introduction to Deep LearningAn Introduction to Deep Learning
An Introduction to Deep Learning
 
Introduction to Deep Learning
Introduction to Deep LearningIntroduction to Deep Learning
Introduction to Deep Learning
 
Using SigOpt to Tune Deep Learning Models with Nervana Cloud
Using SigOpt to Tune Deep Learning Models with Nervana CloudUsing SigOpt to Tune Deep Learning Models with Nervana Cloud
Using SigOpt to Tune Deep Learning Models with Nervana Cloud
 

Recently uploaded

{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...Pooja Nehwal
 
Industrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdfIndustrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdfLars Albertsson
 
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...dajasot375
 
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...ThinkInnovation
 
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfRadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfgstagge
 
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Callshivangimorya083
 
代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改
代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改
代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改atducpo
 
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...Florian Roscheck
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一F sss
 
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...Suhani Kapoor
 
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...Sapana Sha
 
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)jennyeacort
 
(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service
(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service
(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Serviceranjana rawat
 
Data Science Jobs and Salaries Analysis.pptx
Data Science Jobs and Salaries Analysis.pptxData Science Jobs and Salaries Analysis.pptx
Data Science Jobs and Salaries Analysis.pptxFurkanTasci3
 
Dubai Call Girls Wifey O52&786472 Call Girls Dubai
Dubai Call Girls Wifey O52&786472 Call Girls DubaiDubai Call Girls Wifey O52&786472 Call Girls Dubai
Dubai Call Girls Wifey O52&786472 Call Girls Dubaihf8803863
 
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...Suhani Kapoor
 
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /WhatsappsBeautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsappssapnasaifi408
 
Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝soniya singh
 

Recently uploaded (20)

{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...{Pooja:  9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
{Pooja: 9892124323 } Call Girl in Mumbai | Jas Kaur Rate 4500 Free Hotel Del...
 
Industrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdfIndustrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdf
 
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
 
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
Predictive Analysis - Using Insight-informed Data to Determine Factors Drivin...
 
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfRadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdf
 
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
 
代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改
代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改
代办国外大学文凭《原版美国UCLA文凭证书》加州大学洛杉矶分校毕业证制作成绩单修改
 
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
 
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
 
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
Saket, (-DELHI )+91-9654467111-(=)CHEAP Call Girls in Escorts Service Saket C...
 
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
 
(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service
(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service
(PARI) Call Girls Wanowrie ( 7001035870 ) HI-Fi Pune Escorts Service
 
Data Science Jobs and Salaries Analysis.pptx
Data Science Jobs and Salaries Analysis.pptxData Science Jobs and Salaries Analysis.pptx
Data Science Jobs and Salaries Analysis.pptx
 
VIP Call Girls Service Charbagh { Lucknow Call Girls Service 9548273370 } Boo...
VIP Call Girls Service Charbagh { Lucknow Call Girls Service 9548273370 } Boo...VIP Call Girls Service Charbagh { Lucknow Call Girls Service 9548273370 } Boo...
VIP Call Girls Service Charbagh { Lucknow Call Girls Service 9548273370 } Boo...
 
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
 
Dubai Call Girls Wifey O52&786472 Call Girls Dubai
Dubai Call Girls Wifey O52&786472 Call Girls DubaiDubai Call Girls Wifey O52&786472 Call Girls Dubai
Dubai Call Girls Wifey O52&786472 Call Girls Dubai
 
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
 
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /WhatsappsBeautiful Sapna Vip  Call Girls Hauz Khas 9711199012 Call /Whatsapps
Beautiful Sapna Vip Call Girls Hauz Khas 9711199012 Call /Whatsapps
 
Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Defence Colony Delhi 💯Call Us 🔝8264348440🔝
 

Sigma Xi Research Showcase 2018 - Oleksii Volkovskyi

  • 1. Neural Network Hyperparameter Optimization using Q-Learning Oleksii Volkovskyi American Heritage School
  • 2. Neural Network Background ● A statistical model ○ Given Inputs, computes an Output ○ Inspired by the Human Brain ○ Consists computational Units with Parameters ■ Uses complex nonlinear functions in order to achieve high results ○ Relies on computational power and data size ○ Backpropagation can be used to “train” the network ■ Works on the principle of comparing current output to the dataset output and self modification in order to increase accuracy
  • 3. Neural Network Hyperparameters ● Hyperparameters ○ Constants chosen before Training ● Regularisation Constant (Alpha) ○ Throttles complexity of trained model ○ Small constant values results in simpler models ○ Is generally randomly selected ● Size of Hidden Layers ○ Affects computational complexity, and can have a negative effect ○ Is generally a human decision ○ Computational Expense ● Random selection and arbitrary decision can lead to suboptimal performance and failure to reach the global minimum of the neural network (bottom right graph)
  • 4. Q Learning Algorithm ● In order to tackle the problem of suboptimal hyperparameters, a reinforcement learning algorithm can be used ○ Q learning is an algorithm that makes decision given a state ○ The algorithm defines a Q-matrix that acts as a decision matrix ■ The algorithm iteratively improves the Q-matrix given a training dataset ○ Q learning can be used with a dynamic Q-matrix that can be modified when a new state is created, which means that the number of states doesn’t have a limit
  • 5. Q Learning Algorithm ● Q - Learning will be applied to the Neural Network Hyperparameter Selection task as following: ○ The Q-Matrix will be a multidimensional array, depending on the number of hyperparameters and chosen states ■ In this case, the chosen hyperparameter was the regularisation constant ■ The state will be defined as an array of the following: ● The current regularisation constant value ● The current bias-variance metrics ○ The reward will be computed as the F-score of the algorithm (accuracy metric) ○ When trained on multiple networks, the algorithm should be able to generalise to any Neural Network ■ The goal will be to converge to Optimal Hyperparameters in shorter computational time
  • 6. Hypothesis ● Q-learning, with hyperparameters and bias-variance metrics given as states and the F-score as reward, can be used to develop a general algorithm for tuning the hyper- parameters of a neural network. ● General neural network algorithms will: ○ Eliminate the need for human interference in neural network algorithm training ○ Maximise the capabilities of the neural network concept
  • 7. Bias Variance Metrics and F-Score ● Bias and Variance Metrics will be defined by the given formulas ○ They are responsible for the intuition behind Q-learning algorithm ○ Bias - The ability of an algorithm to fit training data ○ Variance - The ability of an algorithm to generalise to test data ○ Given these metrics, the Q-learning system is tasked with scaling the regularisation constant ■ Increasing - Simpler system ■ Decreasing - Complex System ● F-score is a more rigorous alternative to accuracy ○ Handles skewed datasets more strictly ○ Eliminates the need for large pre-processing of data
  • 8. Procedure 1. Create a Neural Network Function, that Inputs Regularisation Alpha, Outputs F-score, Bias, and Variance. 2. Write a function that selects action based on Q-matrix Reward and euclidean distance to possible hyperparameter values 3. Write the Q-learning iterative learning loop, that updates the Q-matrix based on the computed reward and action chosen 4. Train Algorithm on Train Data, which I have chosen to be a poisonous mushroom classification dataset 5. Benchmark algorithm on Test Data a. Record Time and Final Accuracy of Algorithm (Average across 10 Runs) 6. Run Random Selection algorithm on Test Data as Control (Average data from 10 Runs) a. Since random selection can narrow down to an infinitely small range of hyperparameters, the results will compare computational time, while accuracy will be a controlled variable
  • 9. Results ● Q-learning yielded a decrease in Average Computational Time when compared to Random Selection ○ 28.65% on the test dataset ○ 47.05% on the train dataset ● It was given the task to generalise from training to test datasets ○ The datasets had similar concepts, yet entirely different tasks ■ Safe/Poisonous Mushroom Classification (Train) ■ Malignant/Benign Cancer Classification (Test) ○ The algorithm was prone to overfitting on the train dataset ■ The algorithm was trained on one dataset due to limitations in hardware and time
  • 10. Conclusion ● The hypothesis was validated ○ The Q-Learning algorithm was able to fit hyperparameters when given an unknown dataset ■ Additionally, the Q-Learning algorithm Outperformed baseline algorithm (Random Selection) in terms of computational time ○ The improvements in computational time were made due to proper selection of hyperparameters, without modifying the core of the neural network
  • 11. Future Improvements ○ Large improvements can be gained from more complex reinforcement learning algorithms ■ One example is Deep Q-Learning, used by Google in the development of AlphaGo ○ Higher variety of Hyperparameters would validate the experiment further ■ Hyperparameters such as hidden layer size and amount of layers could be used, as well as the choice of computational unit function ■ This would come with an increase in training time and overall computational time, because the algorithm would have to consider exponentially more options ○ Training on a variety of datasets would eliminate overfitting
  • 12. Applications ● General neural network algorithms would: ○ Eliminate the need for human interference ○ Maximise Performance of neural networks ○ Have a chance to reach Bayes’ Optimal Error, which is defined as the smallest error possible for a function to have on a dataset ● Hyperparameter training is applicable to other, more complex, types of Neural Networks ○ Convolutional Neural Networks ■ Autonomous Driving ■ Face Recognition / Verification ○ Recurrent Neural Networks ■ Speech Recognition ■ Music Composition