COMP24111: Machine Learning and Optimisation Chapter 1A: Machine - PowerPoint PPT Presentation

COMP24111: Machine Learning and Optimisation Chapter 1A: Machine Learning Basics Dr. Tingting Mu Email: tingting.mu@manchester.ac.uk

Machine Learning • A machine learning system is a magic box that can be used to – Automate a process – Automate decision making – Extract knowledge from data – Predict future event – Adapt systems dynamically to enable better user experiences – … • How do we build a machine learning system? 1

Machine Learning • “The goal of machine learning is to make a computer learn just like a baby — it should get better at tasks with experience.” • Basic idea: – To represent experiences with data . – To covert a task to a parametric model . – To convert the learning quality to an objective function . – To determine the model through optimising an objective function. • Machine learning research builds on optimisaton theory, linear algebra, statistics … 2

Example: Wine Classification • Wine experts identify the grape type by smelling and tasting the wine. • The chemist says that wines derived from different grape types are different in terms of alcohol, malic acid, alcalinity of ash, magnesium, color intensity, etc. • We get the measurements. But, too many numbers … Can build a machine learning system to automate grape type iden5fica5on! 3

Example: Wine Classification • Task: To identify the grape type of a wine sample based on its chemical quantities measured! Feature v Collecting wine samples for each grape type. Extrac5on v Characterising each wine sample with 13 chemical features. Experiences feature vectors class labels x 1 = x 1,1 , x 1,2 , x 1,3 , … , x 1,12 , x 1,13 ⎡ ⎤ ⎦ , y 1 = grape type 1 ⎣ x 2 = x 2,1 , x 2,2 , x 2,3 , … , x 2,12 , x 2,13 ⎡ ⎤ ⎦ , y 2 = grape type 2 ⎣ 30 bo=les in total, 10 bo=les for each tree type, x 3 = x 3,1 , x 3,2 , x 3,3 , … , x 3,12 , x 3,13 ⎡ ⎤ ⎦ , y 3 = grape type 2 ⎣ each bo=le is charaterised with 13 features. ! ! x 30 = x 30,1 , x 30,2 , x 30,3 , … , x 30,12 , x 30,13 ⎡ ⎤ ⎦ , y 30 = grape type 1 ⎣ 4

Example: Wine Classification v Design a mathematical model to predict the grape type. The model below is controlled by 14 parameters: w 1 , w 2 , … , w 13 , b [ ] Wine features. Predicted grape type by computer. ⎧ 13 ✔ y 1 ⎪ ∑ type 1, if w i x i + b ≥ 0 ⇒ ˆ bottle 1: x 1 y 1 = g ( x 1 ) ⎪ ✗ y 2 i = 1 ˆ ⇒ ˆ y = g ( x ) = y 2 = g ( x 2 ) ⎨ bottle 2: x 2 ! 13 ⎪ ! ! ! ∑ type 2, if w i x i + b < 0 ⎪ ✔ y 30 ⇒ ˆ y 30 = g ( x 30 ) bottle 30: x 30 ⎩ i = 1 Real grape type. v System training is the process of finding the best model parameters by minimising a loss function. ⎡ * , … , w 13 ⎤ O loss w 1 , w 2 , … , w 13 , b ( ) * , w 2 * , b * w 1 ⎦ = argmin ⎣ w 1 , w 2 , … , w 13 , b Loss: predic5ve inaccuracy 5

Example: Wine Classification • Now, given an unseen bottle of wine: ⎧ 13 + b * ≥ 0 ⎪ ∑ w * type 1, if i x i ⎪ i = 1 ˆ y = g ( x ) = ⎨ 13 ⎪ + b * < 0 ∑ w * type 2, if i x i ⎪ ⎩ i = 1 13 Features: x 1 = 12.25, x 2 = 3.88, x 3 = 2.2, x 4 = 18.5, x 5 = 112, x 6 = 1.38, x 7 = 0.78, x 8 = 0.29, x 9 = 1.14, x 10 = 8.21, x 11 = 0.65, x 12 = 2, x 13 = 855 6

The World Generates Data! • Data is recorded on real-world phenomenons. The World is driven by data. – Germany’s climate research centre generates 10 petabytes per year. – Google processes 24 petabytes per day. – PC users crossed over 300 billion videos in August 2014 alone, with an average of 202 videos and 952 minutes per viewer. – There were 223 million credit card purchases in March 2016, with a total value of £12.6 billion in UK. – Photo uploads in Facebook is around 300 million per day. – Approximately 2.5 million new scientific papers are published each year. – … • What might we want to do with that data? – Prediction - what can we predict about this phenomenon? – Description - how can we describe/understand this phenomenon in a new way? • Humans cannot handle manually data in such scale any more. A machine learning system can learn from data and offer insights. 7

Machine learning is important! Speech Recogni5on Speech All of these Robo5cs Synthesis are subfields of Artificial Intelligence (A.I.) Machine Learning Data Natural Mining, Language Analysis, Processing Engineering Computer Text Mining Vision 8

COMP24111, Machine Learning and School Courses Optimisation COMP61011, Foundations of Machine Learning COMP14112, COMP61021, Modelling and visualization Fundamentals of A.I. of high-dimensional data Speech Recogni5on COMP34120, AI and Games Speech Robo5cs Synthesis Machine COMP60711, Learning Data Engineering Data Natural Mining, Language Analysis, Processing Engineering COMP38120, Documents, Computer Text Mining Services and Data on the Web COMP37212, Vision Computer Vision COMP61332, Text Mining 9

Learning Type: Supervised • In supervised learning, there is a “teacher” who provides a target output for each data pattern. This guides the computer to build a predictive relationship between the data pattern and the target output. • The target output can be a real-valued number, an integer, a symbol, a set of real-valued numbers, a set of integers, or a set of symbols. • A training example (or called sample) is a pair consisting of an input data pattern (or called object) and a target output. • A test example is used to assess the strength and utility of a predictive relationship. Its target output is only used for evaluation purpose, and never contributes to the learning process. • Typical supervised learning tasks include classification and regression . 10

Classification Examples: The target output is a category label. • Medical diagnosis: x=patient data, y=positive/negative of some pathology • Optical character recognition: x=pixel values and writing curves, y=‘A’, ‘B’, ‘C’, … • Image analysis: x=image pixel features, y=scene/objects contained in image • Weather: x=current & previous conditions per location, y=tomorrow’s weather … … … this list can never end, applications of classification are vast and extremely active! 11

Regression Examples: The target output is a continuous number (or a set of such numbers). • Finance: x=current market conditions and other possible side information, y=tomorrow’s stock market price • Social Media: x=videos the viewer is watching on YouTube, y=viewer’s age • Robotics: x=control signals sent to motors, y=the 3D location of a robot arm end effector • Medical Health: x=a number of clinical measurements, y=the amount of prostate specific antigen in the body • Environment: x=weather data, time, door sensors, etc., y=the temperature at any location inside a building … … … this list can never end, applications of regression are vast and extremely active! 12

Successful Applications • Convert speech to text, translate from one language to the other. 13

Successful Applications • Face recognition 14

Successful Applications • Object recognition, speech synthesis, information retrieval. 15

Learning Type: Unsupervised • In unsupervised learning, there is no explicit “teacher”. • The systems form a natural “understanding” of the hidden structure from unlabelled data. • Typical unsupervised learning task includes – Clustering: group similar data patterns together. – Generative modelling: estimate distribution of the observed data patterns. – Unsupervised representation learning: remove noise, capture data statistics, capture inherent data structure. MATLAB’s example MATLAB’s example 16 From https://cambridge-intelligence.com/keylines-network-clustering/

Successful Applications • Document clustering and visualisation 17

Learning Type: Reinforcement • In reinforcement learning, there is a “teacher” who provides feedback on the action of an agent, in terms of reward and punishment. • Examples: – Helicopter manoeuvres: reward for following desired trajectory, punishment for crashing – Manage an investment portfolio: reward for each $ in bank – Control a power station: reward for producing power, punishment for exceeding safety thresholds – Make a humanoid robot walk, reward for forward motion, punishment for falling over – Play many different Atari games better than humans: reward for increasing score, punishment for decreasing score These examples are from UCL course on RL. 18

Successful Applications • Game player, self-driving cars, trading strategy. 19

History • 1940s, Human reasoning / logic first studied as a formal subject within mathematics (Claude Shannon, Kurt Godel et al). • 1950s, The Turing Test is proposed: a test for true machine intelligence, expected to be passed by year 2000. Various game-playing programs built. 1956, Dartmouth conference coins the phrase artificial intelligence . 1959, Arthur Samuel wrote a program that learnt to play draughts (checkers if you are American). • 1960s, A.I. funding increased (mainly military). Famous quote: Within a generation ... the problem of creating 'artificial intelligence' will substantially be solved." • 1970s, A.I. winter. Funding dries up as people realise it is hard. Limited computing power and dead-end frameworks. 20