MACHINE LEARNING & INDUSTRY

Review and trends

Dr. Esteban Guerrero

esteban@cs.umu.se

AGENDA

1. Introduction
2. Machine Learning and Industry, European and Swedish context
3. ML approaches (hands-on)
4. Industry 4.0
5. ML in two industrial sectors (cases)
6. Concluding remarks
7. Activity: co-analysis of ML+IND scenarios

DR. ESTEBAN GUERRERO

• Current position: Researcher
• Industrial experience:
• 4 years R&D engineer in a Colombian telecommunications company. Role: support back-end services.
• Education:
• Ph.D. in Computing Science, Umeå University. Sweden.
• Ph. Licentiate in Computing Science, Umeå University.
• M.Sc. Master’s degree in Computer Science, Malmö University. Sweden.
• M.Sc. Master’s studies in Telematics Engineering, University of Cauca. Colombia.
• B.Eng. Bachelor degree in Electronic and Telecommunications Engineering, University of Cauca. Colombia.

1. Introduction

https://people.cs.umu.se/esteban/

1. Introduction

SmartVib

en liten separat enhet med inbyggda sensorer som trådlöst sänder signaler till mobiltelefonen i vilken en applikation behandlar de uppmätta vibrationerna.

EUROPEAN CONTEXT

Main source: https://ec.europa.eu/info/index_sv

https://ec.europa.eu/info/sites/info/files/research_and_innovation/knowledge_publications_tools_and_data/documents/ec_rtd_radical-innovation-breakthrough_052019.pdf

2. Context

Key in the report:

• Summary of “recent progress directions” in artificial intelligence:
• Duelling networks, also called generative adversarial networks (GANs)
• Capsule Networks
• Long term perspectives:

“In the next decade and beyond, one can expect significant research around one-shot and zero-shot learning models involving knowledge transfer.”

https://ec.europa.eu/info/sites/info/files/research_and_innovation/knowledge_publications_tools_and_data/documents/ec_rtd_radical-innovation-breakthrough_052019.pdf

2. Context

2. Context

2. Context

2. Context

AI techniques

2. Context

AI functional applications

2. Context

AI functional applications

PATENTS - GLOBAL

Additional info:

• G‑II, 3.3.1 Artificial intelligence and machine learning - Guidelines for Examination (epo.org)

WIPO Technology Trends 2019: Artificial Intelligence

PATENTS - SWEDEN

Additional info:

• https://www.prv.se/
• Documentation in Canvas

Antal ansökningar per detaljområde 2010–2016

De tre huvudområdena låter sig delas in i underliggande detaljområden.

Patent- och registreringsverkets statistikårsbok 2018 (prv.se)

Under perioden 2010–2016 har antalet PCT-ansökningar inom Industri från svenska sökanden 4.0 en stark tillväxt. Faktum är att antalstillväxten i tidsperioden är dubbelt så stor för svenska ansökningar som för europeiska.

SWEDISH CONTEXT

Main sources: Vinnova, Regeringskansliet 

2. Context

SMART INDUSTRY - A STRATEGY FOR NEW INDUSTRIALISATION FOR SWEDEN�

Four focus areas have been chosen:

• Industry 4.0 – Exploit the potential of digitalisation
• Sustainable production – Improve the industrial sector’s capacity for sustainable and resource-efficient production
• Industrial skills boost – Ensure the supply of skills to the industrial sector
• Test bed Sweden – Create attractive innovation environments

https://www.government.se/information-material/2016/04/smart-industry---a-strategy-for-new-industrialisation-for-sweden/

https://www.regeringen.se/

2. Context

https://www.government.se/49e94a/contentassets/3be3b6421c034b038dae4a7ad75f2f54/nist_a4_faktablad_160701_eng_web.pdf

2. Context

Svenska AI-miljöer med geografisk hemvist.

39 AI-miljöer som arbetar för utveckling av artifciell intelligens

Centers with focus

on Machine Learning

https://www.vinnova.se/globalassets/mikrosajter/ai/vr_19-05_190704.pdf

2. Context

INDUSTRY 4.0

INDUSTRIE 4.0

The Fourth Industrial Revolution aims to leverage differences between the physical, digital, and biological sphere. It integrates cyber-physical systems and the Internet of Things, big data and cloud computing, robotics, artificial-intelligence based systems and additive manufacturing

Expected effects:

• On the business side: it drastically modifies customer expectations, product enhancement, collaborative innovation and organisational forms. 
• On people: one of the greatest challenges is on privacy, on the notion of ownership, consumer patterns and how we devote time to develop skills.

European Commission, strategy on digitising the European industry. 

https://ec.europa.eu/digital-single-market/en/fourth-industrial-revolution

4. Industry 4.0

INDUSTRIE 4.0

4. Industry 4.0

Frank, A. G., Dalenogare, L. S., & Ayala, N. F. (2019). Industry 4.0 technologies: Implementation patterns in manufacturing companies. International Journal of Production Economics, 210, 15-26.

Industrial Machine Learning (at GE)

Joshua Bloom, Professor at UC Berkeley, CTO Wise/GE

Industrial Machine Learning (at GE)

Joshua Bloom, Professor at UC Berkeley, CTO Wise/GE

https://www.mmcventures.com/wp-content/uploads/2019/02/The-State-of-AI-2019-Divergence.pdf

4. Industry 4.0

Leitão, P., Colombo, A. W., & Karnouskos, S. (2016). Industrial automation based on cyber-physical systems technologies: Prototype implementations and challenges. Computers in Industry, 81, 11-25.

Challenges in industrial cyber-physical systems�

4. Industry 4.0

ML APPROACHES (BRIEF)�+ �HANDS-ON

MACHINE LEARNING LANGUAGES

Swift

3. ML approaches

SUPERVISED LEARNING

• What is: “Supervised learning is a learning model built to make prediction, given an unforeseen input instance.”

​

• How it works: “With supervised learning you use labelled data, which is a data set that has been classified, to infer a learning algorithm. The data set is used as the basis for predicting the classification of other unlabelled data” [Talabis,et.al.2014].

​

Talabis, M., McPherson, R., Miyamoto, I., & Martin, J. (2014). Information Security Analytics: Finding Security Insights, Patterns, and Anomalies in Big Data. Syngress.

3. ML approaches

SUPERVISED LEARNING

• Two important approaches (among many):
• Regression: the goal is to predict a continuous measurement for an observation. That is, the responses variables are real numbers. Applications include forecasting stock prices, energy consumption, or disease incidence.
• Classification: the goal is to assign a class (or label) from a finite set of classes to an observation. That is, responses are categorical variables. Applications include spam filters, advertisement recommendation systems, and image and speech recognition.

3. ML approaches

SUPERVISED LEARNING

Example 0: Linear regression, a diabetes dataset with Python

Basic concepts

• The input or predictor variable is the variable(s) that help predict the value of the output variable. It is commonly referred to as X.
• The output variable is the variable that we want to predict. It is commonly referred to as Y.
• We assume the equation: Yₑ = ε + β X
• Yₑ is the estimated or predicted value of Y based on our linear equation.

Goal: find statistically significant values of the parameters ε and β that minimize the difference between Y and Yₑ.

If we are able to determine the optimum values of these two parameters, then we will have the line of best fit that we can use to predict the values of Y, given the value of X.

3. ML approaches

SUPERVISED LEARNING

• Example 0: Linear regression, a diabetes dataset with Python
• Jupyter notebook using Python3, scikit-learn

https://colab.research.google.com/drive/1Ey0_E-fCtggCNxdd96v8gxbCnAKh7_ug

​

3. ML approaches

https://www4.stat.ncsu.edu/~boos/var.select/diabetes.tab.txt

https://web.stanford.edu/~hastie/Papers/LARS/LeastAngle_2002.pdf

SUPERVISED LEARNING

What the line means? 🡪 it represents a function

3. ML approaches

f

Patient x: age 35, sex 1,…

Output: diabetic 82%

UNSUPERVISED LEARNING

• What is: “Unsupervised learning finds structures in the data.”

​

• How does it work?: “Labels for the data instances or other forms of guidance for training are not necessary. This makes unsupervised learning attractive in applications where data is cheap to obtain, but labels are either expensive or not available.” [Wittek. 2014].

​

Wittek, P. (2014). Quantum machine learning: what quantum computing means to data mining. Academic Press.

3. ML approaches

UNSUPERVISED LEARNING

• Two important approaches (among many):
• Clustering (many sub-mechanisms)
• K-Means
• Hierarchical clustering
• Many others
• Principal Components Analysis
• used to preprocess and reduce the dimensionality of high-dimensional datasets, preserving the original structure

​

Wittek, P. (2014). Quantum machine learning: what quantum computing means to data mining. Academic Press.

3. ML approaches

UNSUPERVISED LEARNING

Example 1: K-means with Python + Scikit-learn

K-means algorithm.

Goal: formation of stable clusters

It starts by randomly choosing a centroid value for each cluster. After that the algorithm iteratively performs three steps:

1. Find the Euclidean distance between each data instance and centroids of all the clusters;
2. Assign the data instances to the cluster of the centroid with nearest distance;
3. Calculate new centroid values based on the mean values of the coordinates of all the data instances from the corresponding cluster.

Iris flower data set clustering

3. ML approaches

UNSUPERVISED LEARNING

Example 1: K-means with Python + Scikit-learn

Iris flower data set clustering �

​

​

​

​

https://scikit-learn.org/stable/auto_examples/cluster/plot_cluster_iris.html#sphx-glr-auto-examples-cluster-plot-cluster-iris-py

3. ML approaches

UNSUPERVISED LEARNING

Example 1: K-means with Python + Scikit-learn

Jupyter notebook:

https://scikit-learn.org/stable/auto_examples/cluster/plot_cluster_iris.html#sphx-glr-auto-examples-cluster-plot-cluster-iris-py

https://colab.research.google.com/drive/1dCb2q2ONQA3geg2VBZ03bvxSdmGV_JF0

3. ML approaches

DEEP LEARNING

3. ML approaches

DEEP LEARNING

3. ML approaches

DEEP LEARNING

Example 2: Fashion MNIST dataset + TensorFlow + Python

Goal: Classify images.

How: Train a network with 60000 examples. Evaluate with 10000.

Reference: https://www.tensorflow.org/tensorboard/tensorboard_in_notebooks

3. ML approaches

DEEP LEARNING

Example 2: Fashion MNIST dataset + TensorFlow + Python

​

See Jupyter notebook:

https://colab.research.google.com/drive/13YG9iSbbDRDKPOrD4sE0G8w_Vd2Q-_sF

Reference: https://www.tensorflow.org/tensorboard/tensorboard_in_notebooks

3. ML approaches

REINFORCEMENT LEARNING

• Basic idea:
• Agent receives feedback in the form of rewards
• Agent’s utility is defined by the reward function
• Must (learn to) act so as to maximize expected rewards
• All learning is based on observed samples of outcomes

3. ML approaches

https://commons.wikimedia.org/wiki/File:Reinforcement_learning_diagram.svg

REINFORCEMENT �LEARNING

3. ML approaches

https://deepmind.com/blog/article/deep-reinforcement-learning

https://gym.openai.com/envs/FetchPickAndPlace-v0/

REINFORCEMENT LEARNING

3. ML approaches

REINFORCEMENT LEARNING

3. ML approaches

https://en.wikipedia.org/wiki/File:Markov_Decision_Process.svg

..and we know that the action is determined by

Markov Decision

Process

REINFORCEMENT LEARNING

• Example 3:OpenAI Gym

https://gym.openai.com/

Lubuntu

Python 3.7

3. ML approaches

REVIEW OF �ML IN TWO SPECIFIC SECTORS

Oil and Gas 4.0

Cybersecurity

5. Cases

SPECIFIC INDUSTRY AREAS

Case: Cybersecurity (1/3)

• Spam detection: “The detection of spam is based on the use of filters that analyse the content and decide whether or not they are spam or legitimate messages, blogs or websites” Two main strategies can be followed to detect spam: 1) textual analysis and 2) image-based analysis. �
• Major technologies: Bayesian classifiers: Naive Bayes classifiers, Boolean Naive Bayes, etc.; Support vector machines (SVM), back-propagation neural networks, among others, Deep Belief Networks �

Torres, J. M., Comesaña, C. I., & García-Nieto, P. J. (2019). Machine learning techniques applied to cybersecurity. International Journal of Machine Learning and Cybernetics, 1-14.

Berman, D. S., Buczak, A. L., Chavis, J. S., & Corbett, C. L. (2019). A survey of deep learning methods for cyber security. Information, 10(4), 122.

5. Cases

SPECIFIC INDUSTRY AREAS

Case: Cybersecurity (2/3)

• Malware detection: “The detection of spam is based on the use of filters that analyse the content and decide whether or not they are spam or legitimate messages, blogs or websites” Two main strategies can be followed to detect spam: 1) textual analysis and 2) image-based analysis. �
• Main technologies: Bayesian classifiers: Naive Bayes classifiers, Boolean Naive Bayes, etc.; Support vector machines (SVM), convolutional neural networks (CNNs) and recurrent neural networks (RNNs)

​

​

Torres, J. M., Comesaña, C. I., & García-Nieto, P. J. (2019). Machine learning techniques applied to cybersecurity. International Journal of Machine Learning and Cybernetics, 1-14.

Berman, D. S., Buczak, A. L., Chavis, J. S., & Corbett, C. L. (2019). A survey of deep learning methods for cyber security. Information, 10(4), 122.

5. Cases

SPECIFIC INDUSTRY AREAS

Case: Cybersecurity (3/3)

• Phishing detection: “The detection of spam is based on the use of filters that analyse the content and decide whether or not they are spam or legitimate messages, blogs or websites” Two main strategies can be followed to detect spam: 1) textual analysis and 2) image-based analysis. �
• Main technologies: Bayesian classifiers, SVMs, neural networks

Torres, J. M., Comesaña, C. I., & García-Nieto, P. J. (2019). Machine learning techniques applied to cybersecurity. International Journal of Machine Learning and Cybernetics, 1-14.

Berman, D. S., Buczak, A. L., Chavis, J. S., & Corbett, C. L. (2019). A survey of deep learning methods for cyber security. Information, 10(4), 122.

5. Cases

SPECIFIC INDUSTRY AREAS

Datasets for cybersecurity:

• Knowledge Discovery and Dissemination (KDD) 1999 dataset: https://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html
• NSL-KDD dataset: https://www.unb.ca/cic/datasets/nsl.html
• CTU-13 Dataset: https://www.stratosphereips.org/datasets-ctu13/
• Others: Contagio, Comodo, the Genome Project, Virus Share, VirusTotal, DREBIN, and Microsoft Malware Classification

Torres, J. M., Comesaña, C. I., & García-Nieto, P. J. (2019). Machine learning techniques applied to cybersecurity. International Journal of Machine Learning and Cybernetics, 1-14.

5. Cases

SPECIFIC INDUSTRY AREAS

• Case: Oil and Gas

​

SPECIFIC INDUSTRY AREAS

• Case: Oil and Gas

​

Lu, H., Guo, L., Azimi, M., & Huang, K. (2019). Oil and Gas 4.0 era: A systematic review and outlook. Computers in Industry, 111, 68-90.

5. Cases

GAMING INDUSTRY

• See AI approaches for Gaming (Oct 2019 Unite Copenhagen)

https://www.youtube.com/watch?v=2M3ytOo7LQQ&feature=youtu.be

5. Cases

CONCLUDING

WHY DO BUSINESSES FAIL AT MACHINE LEARNING?�CHECKLIST

• Know what business you’re in.
• Do things in the right order. (don’t start with the algorithms; solve: what business problem I am solving?)
• Don’t reinvent the wheel
• Data is not the most important part (data not Data)
• To scale using ML in Industry, more humans are needed (more than data scientists / engineers; from: statistics, ethics, social work, etc.)

​

https://www.youtube.com/watch?v=dRJGyhS6gA0

Cassie Kozyrkov

Head of Decision Intelligence, Google.

• Simplify where is possible
• Focus more in information data than the algorithms
• Design incentives that cannot be gamed (for reward functions in Reinforcement Learning: what you want the thing to learn?; and maybe for some Deep Learning approaches)

​

​

​

​

6. Conclusions

https://www.ai-playbook.com/introduction

6. Conclusions

END

Questions?

ACTIVITY

1. Individually or in group
2. Choose an industrial use case (it can be anonymized or hypothetical)
3. Simplify the case to I/ML/O
1. Output: ? (key: think in business model)
2. Input: data, human resources, infrastructure, etc.
2. ML approach?
5. Take some minutes to reflect.
6. Group analysis

​

7. Activity

PROGRAMMING WITH PYTHON

https://www8.cs.umu.se/kursmaterial/ws-python-scientific-computing/

https://www.mmcventures.com/wp-content/uploads/2019/02/The-State-of-AI-2019-Divergence.pdf

NEW REPORT (OCTOBER)

http://sip-piia.se/wp-content/uploads/2019/10/AI-rapport_2019-low.pdf