IN5490 Advanced Topics in Artificial

Intelligence for Intelligent Systems

Lecture 1 – 2018 Course Introduction

Jim Tørresen

• Lecturer:

– Bruno Castro da Silva – bruno.silva@inf.ufrgs.br (UFRGS)

– Charles Martin – charlepm@ifi.uio.no (UiO)

– Enrique Garcia Ceja – enriqug@ifi.uio.no (UiO)

– Zia Uddin – mdzu@ifi.uio.no (UiO)

– Jim Tørresen – jimtoer@ifi.uio.no (UiO)

– and more

• Lecture time: 27–31 August, 15-19 October, 19-23 November

• Lecture room: ROBIN pause area

• Course web page:

https://www.uio.no/studier/emner/matnat/ifi/IN5490/index.html2

IN5490 Advanced Topics in Artificial

Intelligence for Intelligent SystemsAutumn 2018

Topics to be covered in the course

• Classification

• Recurrent neural networks (RNN)

• Deep re-inforcement learning (DRL)

• Convolutional Neural Networks (CNN)

applied to sensor data analysis

• Mobile robotics

3

Learning outcome

• insight into the new and promising methods used in

artificial intelligence (AI) and machine learning (ML)

• have knowledge about how to apply AI methods to

different kinds of applications

• be able to search for literature outlining state-of-the-

art within a specific research field.

• be able to critically assess scientific papers and be

familiar with how to prepare a scientific paper

• be able to design and conduct experiments using AI

methods, with emphasis on evaluation

• have experience in presenting scientific work 4

Course Layout

• Teaching:

– Lectures and organized group work in selected weeks

– Self study and group work in between sessions

• Grading: Pass/not-pass

• To pass:

– give at least one presentation (one scientific paper)

– prepare one research paper draft within each

project group (will be considered to be submitted to

a conference)

– attend at least 80% of all seminar sessions5

6

9th Joint IEEE International Conference on

Development and Learning and on Epigenetic Robotics

19-22 August 2019, Oslo, Norway

Temp. web page: https://icdlepirob2019.wordpress.com

Call for Papers/Workshop/Tutorials flyer available

Paper review and presentation

• Pick one from our list of papers (add your

name to it, only one student can present a

paper)• https://docs.google.com/spreadsheets/d/1xfIJZyOcUallpB8SF1

4JkpkdNKisC1oa5FSUPpjxwJI/edit#gid=0

• Read and understand the paper content

• Prepare a presentation with:

– Paper title page heading

– Main motivation and idea of method in the paper

– Main results

– Assessment of strengths and weaknesses

• Give a presentation in 10 minutes (Oct or Nov)

7

Session Plan 27–31 AugustSee message at course web page for more details

• Monday 14:15-18:00 Course intro + lecture on

publishing papers + initial project workshop + pizza

• Tuesday 14:15-18:00 2 hours lecture + 2 hours

project workshop (select and plan project)

• Wednesday 14:15-18:00 2 hours lecture + 2 hours

project workshop (work on project)

• Thursday 14:15-18:00 2 hours lecture + 2 hours

project workshop (work on project)

• Friday 14:15-16:00 2 hours lecture8

Topics for this week

• Lectures

• Form groups of 3 persons

• Select/define a project to work on in the

course

• Start planning/working on the project

9

Project Implementation

• Preferred programming language: Python

• Preferred tool: Keras + Tensorflow, Jupyter

Notebook, OpenAi Gym

• Group size: 3 students

• Project proposals: https://docs.google.com/document/d/14LZlb0hmp7j-

LPWvEwFeZpB0xV4y-z-84MduhicWuzg

• Register group here: https://docs.google.com/spreadsheets/d/1aNMQO7KCxcoRqwXmg2

8ehmKHm-lHpsopwpfHobpmAWw/edit#gid=0

• Register group: Tuesday at the lastest

• Register selected project: Wednesday at the latest 10

Biology

applyprinciples

from nature

Applications

roboticsmusichealth

++

Hardware electronics3D-printing prototyping

Robotics and Intelligent Systems

Robotics and Intelligent Systems group

ROBIN

Web page: Google for ”ROBIN IFI”

Creating systems for demanding run-time environments.

Robotics and

Intelligent Systems

research

12

Prediction/Forecasting Examples

• Nature and society: natural disasters,

pandemics, demography, population

dynamics and meteorology

• Finance: stock market behaviour

• Sports: outcome of sporting events

• Robotics: More effective operation when

human close by

• Music: Real-time music synthesis together

with real musicians13

Interaction and User Behaviour using

Prediction

14

Why: Responsive and user adapted systems

MECS: Multi-sensor Elderly Care Systems Research Council of Norway grant 247697

Funding: FRINATEK

Research Council of Norway

Goal: Create and evaluate multimodal mobile human supportive

systems that are able to sense, learn and predict future events.

INTROMAT: INtroducing personalized

TReatment Of Mental health problems using

Adaptive Technology (2016-2021)

Goal: Increase access to

mental health services

for common mental health

problems by developing

smartphone technology

which can guide patients.http://intromat.no

Project Manager:

Haukeland Univ. Hospital, Bergen

Funding: IKTPLUSS Lighthouse,

Research Council of Norway

EPEC: Prediction and Coordination for

Robots and Interactive MusicResearch Council of Norway grant 240862.

Goal: Design, implement and evaluate multi-sensor

systems that are able to sense, learn and predict

future actions and events.

What is a Prediction?• Output from a trained supervised model

• Confusing in literature:

– With or without representing any state in time

(temporal data)

– Estimate current or future state of a system

(temporal model)

• A more limited definition (informal speech):

Forecasting

18

Temporal and Non-Temporal Data

Temporal data is data that represents a state in

time

• Non-temporal: time-independent data

– Text (a set of words rather than sequence of words)

– Still Images

• Temporal: a consecutive sequence of data

– Text (sequence)

– Audio

– Video

– Music

– Animation

– Human motion

19

Temporal and Non-Temporal Models

• Non-temporal: consider each input

vector independently

– E.g. Feed-forward neural networks

– Enrique demo

• Temporal: consider past and current

input vector (e.g. using memory or

multiple inputs)

– E.g. Recurrent Neural Networks

– Kai demo20

x

h1

h2

h3

y x

h1

h2

h3

y

Time Delay Network

21

Machine Learning for Prediction: Training Temporal and Non-Temporal Models

• Non-temporal Classification/Recognition• Feed forward neural networks• Convolutional neural networks • Random forest classifier • SVM, K-NN

• Temporal Sequences Prediction• Recurrent neural networks (RNN)

• Long short term memory (LSTM)• Hidden Markov Models• Conditional Random Fields • Dynamic Time Warping

Machine Learning for Prediction: Training Temporal and Non-Temporal Models