Introduction to Machine LearningJim Tørresen – e-mail: jimtoer@ifi.uio.noResearch group Robotics and Intelligent SystemsDepartment of Informatics, University of Oslo

What is the Machine Learning about?• Technology that can adapt through learning• Systems that can sense, reason (think)

and/or act/respond• Inspired from biology/nature• Increase intelligence in both single node and

multiple node systems

5

Man/Woman vs Machine – Who are smartest?

• Machines are good at: – number crunching – storing data and searching in data– specific tasks (e.g. control systems in

manufacturing)• Humans are good at:

– sensing (see, hear, smell etc and be able to recognize what we senses)

– general thinking/reasoning – motion control (speaking, walking etc).

7

Major Mechanisms in Nature

• Evolution: Biological systems develop and change during generations.

• Development/growth: By celldivision a multi-cellular organismis developed.

• Learning: Individuals undergo learning through their lifetime.

• Collective behavior: Immune systems, flocks of birds, fishes etc

• Sensing and motion

AI, machine learning, deep learning…

What Methods are Best?

12

Self learning/Machine learning

System to be designed

Data set/ specification

Algorithm

Data Driven Modeling in Machine Learning

14

Artificial Neural Network

15

Processing Element

2018.02.16 18

Types of Machine Learning (1)• Supervised learning: Training data includesdesired outputs. Based on this training set, thealgorithm generalises to respond correctly to allpossible inputs.

• Unsupervised learning: Training data does notinclude desired outputs, instead the algorithmtries to identify similarities between the inputsthat have something in common are categorisedtogether.

19

Classification

AlgorithmModel

Lion

Lion

Forecasting, prediction and abnormality detection

Deep learning algorithms• Feature extraction/creation

• Deep belief networks (DBN)• Autoencoders

• Classification/recognition• Deep neural networks• Convolutional neural networks

(CNN)• Time series/sequences/prediction

• Recurrent neural networks (RNN) • Long short term memory

2018.02.16 23

Types of Machine Learning (2)• Reinforcement learning: The algorithm is toldwhen the answer is wrong, but does not get toldhow to correct it. Algorithm must balanceexploration of the unknown environment withexploitation of immediate rewards to maximize long-term rewards.

• Evolutionary learning: Biological organisms adaptto improve their survival rates and chance of havingoffspring in their environment, using the idea offitness (how good the current solution is).

2018.02.16 25

Search Space

Divide-and-Conquer

26

Machine Learning Toolboxes (we have used)

• MathWorks/MATLAB Statistics and Machine Learning Toolbox

• Scikit-learn (Python)

• TensorFlow (C++/Python, Google team)

• Weka (Java, lower level than Python )

Other options:

• Theano (Python, competitor to TensorFlow)

• Keras: Deep Learning library for Theano and TensorFlow

• Lasagne Build and train neural networks in Theano (alt. to Keras) 27

Learn More (Courses at Univ of Oslo)

• STK2100 - Machine learning and statistical methods for prediction and classification (first time Spring 2017)

• INF4490 - Biologically inspired computing (autumn, ROBIN group responsible)

• INF5860 - Machine Learning for Image Analysis (first time Spring 2017)

28

Machine Learning Application Examples• Computer systems

– Web search– Web shopping– Optimization e.g. the design of physical shapes– Route planning

• Embedded/physical systems– Smartphone user adaptation– Detecting faces/people smiling in cameras– Service robots– Driverless drones, cars and submarines 29

Incr

easi

ng

size

/com

plex

ity

Where is machine learning used nowadays?

● Almost everywhere ○ Detection○ Diagnosis○ Prediction○ Drug discovery○ Personalised medicine○ Medical Imaging○ Genomics○ And many more!

RITMO/Music Relevant Examples

• Sensing: Classification of Error Types in Physiotherapy Exercises (Haakon Drews master project)

• Acting: PheroMusic (Kristian) and RobotJam(Charles)

33

Detecting mistakes made during physiotherapy exercises

34

Intertial measurement unit (IMU)

Detecting mistakes made during physiotherapy exercises• Three general error types chosen:

– amplitude (too large or small movements) – speed (too fast or slow movements) – trajectory (not moving in the right direction).

• Classification: k-nearest neighbor and dynamic time warping

• Error types are classified with high rates ofaccuracy when using template exercises for a given test subject.

• When classifying error types across subjects, the accuracy proves unsatisfactory.

35

(Inter) Active Music

36

Indirect Controlo Use on-body sensors to adapt the

music to the mood of the usero Listen to music that pushes you to

work out hardero Fuse the musical preferences of

multiple users into one song

Direct Controlo Navigate within the songo Control certain instruments (e.g.

keep playing the chorus drumbeat in the verse)

o Change the tempo of the song

Ant Colony Optimization (ACO)• Ants find shortest path to food source from nest.• Ants deposit pheromone along traveled path

which is used by other ants to follow the trail.• This kind of indirect communication via the local

environment is called stigmergy.

37

Direct + Indirect Control Combined

39

PheroMusic app• Control by ant colony

optimization

Reuters

40

RoboJam (Charles)

A recurrent articial neural network to generatesequences of touchscreen interactions and absolutetimings, rather than high-level musical notes.

41

RoboJam (Charles)

42

predict location of touches

predict time deltas in between events