10. Machine Learning

Size: px

Start display at page:

Download "10. Machine Learning"

Hillary Parsons
5 years ago
Views:

1 Artificial Intelligence 10. Machine Learning Prof. Bojana Dalbelo Bašić Assoc. Prof. Jan Šnajder University of Zagreb Faculty of Electrical Engineering and Computing Academic Year 2016/2017 Creative Commons Attribution NonCommercial NoDerivs 3.0 DRAFT 0.3 Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

2 Outline 1 Intro to Machine Learning 2 Approaches to machine learning 3 Supervised learning 4 Algo 1: Naïve Bayes Classifier 5 Algo 2: Decision Trees Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

3 Data Mining Cup 2010 Using the existing characteristics of a customer s initial order, such as order quantity per type of goods, title and delivery weight, a decision must be made on whether to send a voucher worth EUR The customers who receive a voucher should be those who would not have decided to re-order by themselves. IEEE ICDM 2010 DM Competition Modeling the process of traffic jams formation during morning peak in the presence of roadworks, based on initial information about jams broadcast by radio stations. Input data contain identifiers of road segments closed due to roadworks, accompanied by a sequence of segments where the first jams occurred. The algorithm should predict a sequence of segments where next jams will occur in the nearest future. Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

4 Outline 1 Intro to Machine Learning 2 Approaches to machine learning 3 Supervised learning 4 Algo 1: Naïve Bayes Classifier 5 Algo 2: Decision Trees Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

What is machine learning? Machine learning (Alpaydin 2009) Machine learning is programming computers to optimize a performance criterion using example data or past experience.

5 What is machine learning? Machine learning (Alpaydin 2009) Machine learning is programming computers to optimize a performance criterion using example data or past experience. We have a model defined up to some parameter, and learning is the execution of a computer algorithm to optimize the parameters of the model using the training data or past experience. Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

6 Prediction Based on seen data, the model should be able to predict properties of new, unseen data Otherwise said, the model should be able to generalize From a logical perspective, what the model does is induction: based on a limited sample of data, the model infers a general rule how the data behave Aim of machine learning: build models that generalize well! Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

7 Generalization requires some data! Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

8 Why machine learning? At least three reasons: 1 Problems that are too hard to be solved algorithmically we have no idea how to solve them, because we don t really know how we do it ourselves (e.g., speech recognition) problems that cannot be solved in a traditional (algorithmic) manner AI-complete problems 2 Dynamically adaptable systems e.g., robots, adaptable user interfaces 3 Vast amounts of data vast data is out there, can we tap on the knowledge that it contains? data science, big data, data mining NB: Not all problems deserve an ML solution (e.g., payroll accounting software). Do not use a cannon to kill a fly. Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

9 Case 1: Hard problems Pattern recognition Computer vision Natural language understanding Speech recognition Robotics... Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

10 Example: Recognizing handwritten digits The system is presented patterns (digitalized input) It needs to recognize to which of the predefined 10 classes the input pattern belongs to This is called pattern recognition or pattern classification Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

11 Example: Recognizing handwritten digits The problem is that no one writes perfect digits: there is a large number of variants of each digit Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

12 Example: Recognizing handwritten digits Naive approach: store all possible variants of each digit in memory when presented with a digit (8 8 binary matrix), look it up Q: Why is this not a good idea? A: Because there are too many different patterns: 2 64 = Memory issue assume each pattern is 8 bytes 2 (64+3) bytes = 157 exabytes 2 Time issue assume comparing patterns takes 100 microseconds average look-up time is seconds = 29 million years 3 Conceptual issue even if we could solve the above issues, how do we gather the data in the first place? Manual labeling of each pattern would obviously take much longer than 29 million years! Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

13 Case 2: Dynamically adaptable systems An intelligent systems should be capable of adapting to its environment be capable of learning. If it is capable of learning, it can plan how to act in new situations. Robotics Multi-agent systems Intelligent user interfaces Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

that explain the data and allow us to make inferences/predictions Example: Purchase transactions explain users behavior People who

14 Case 3: Knowledge from data From data to information to knowledge to action Data exists in abundance (web, text, video, audio, experimental data, data warehouses, click data, deep web) On the other hand, knowledge is scarce and expensive Goal: build models that explain the data and allow us to make inferences/predictions Example: Purchase transactions explain users behavior People who bought The da Vinci Code also bought The Five People You Meet in Heaven Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

15 Data Mining Data mining or knowledge discovery in data applies machine learning algorithms to structured and unstructured databases Commerce: market basket analysis, CRM Finances: credit risk assessment, credit card fraud Manufacturing: optimization, troubleshooting Medicine: diagnosis Telecommunications: service optimization, churn prediction Bioinformatics: gene expression analysis, sequence alignment Text mining: document classification, information extraction... Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Data Science http://berkeleysciencereview.

16 Data Science Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

17 Data Science Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

18 The promises and perils of Data Science Most of the knowledge in the world in the future is going to be extracted by machines and will reside in machines. Yann LeCun, Director of AI Research, Facebook If we used all our technology resources, we could actually give people personalized recommendations for every step of your life. Aneesh Chopra, former CTO of the U.S. Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Machine learning and cognitive science Cognitively plausible models Understanding machine learning algorithms Understanding human learning abilities Thought Reading Experiment: http://www.cs.cmu.

19 Machine learning and cognitive science Cognitively plausible models Understanding machine learning algorithms Understanding human learning abilities Thought Reading Experiment: Functional magnetic resonance (fmri) Detects blood flow changes in the brain: active brain areas consume more oxygen Exploits the fact that molecules in the blood react to the magnetic field depending on the level of oxygen Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

20 Deep learning: The current big thing in ML Combines cleverly engineered artificial neural networks and large amounts of data We have had neural networks before, but today we also have: Much faster computer architectures (GPU SIMD) Much more data Reaches state-of-the-art results on a broad range of machine learning tasks Google, Baidu, Paypal, NVIDIA, Facebook,... Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Deep learning: autonomous vehicles (NVIDIA) http://blogs.nvidia.

21 Deep learning: autonomous vehicles (NVIDIA) Various sensors gather large amounts of data which a deep learning-based system uses to solve challenging tasks such as detecting pedestrians, automatic parking, changing lanes,... Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

22 Deep learning: autonomous vehicles (NVIDIA) A classifier recognizes various types of vehicles and traffic signs, even in cases that would be difficult for humans (bad weather conditions etc.) Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Deep dish learning (Google) http://www.popsci.

23 Deep dish learning (Google) Based on a still photo of a dish, the system estimates how many calories it has Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

24 Outline 1 Intro to Machine Learning 2 Approaches to machine learning 3 Supervised learning 4 Algo 1: Naïve Bayes Classifier 5 Algo 2: Decision Trees Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

25 Approaches (1) Supervised learning Data is labeled and comes in pairs (input, output)=(x, y). We need to find a mapping f(x) = y If y is a discrete label: classification If y is a number: regression (2) Unsupervised learning Data comes without labels. We need to find regularities in the data. clustering novelty/outlier detection dimensionality reduction (3) Reinforcement learning Learning an optimal strategy based on trials with a delayed credit. Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Classification example: Predicting election outcomes Election Analytics (http://electionanalytics.cs.illinois.

26 Classification example: Predicting election outcomes Election Analytics ( Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Regression example: predicting NO 2 level emissions Exposure models for traffic related NO2 (http://www.sciencedirect.

27 Regression example: predicting NO 2 level emissions Exposure models for traffic related NO2 ( Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

expressivity similar functionality) Dalbelo Bašić,

28 Clustering example: grouping DNA microarrays Goal: grouping genes with similar expressivity (similar expressivity similar functionality) Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Reinforcement learning: AlphaGo Two deep neural networks (DNN) trained to predict the next move and reduce the search space Reinforcement learning on top of DNN, to learn a playing strategy, trained

29 Reinforcement learning: AlphaGo Two deep neural networks (DNN) trained to predict the next move and reduce the search space Reinforcement learning on top of DNN, to learn a playing strategy, trained by playing parties against itself Uses 40 search threads, 48 CPUs, and 8 GPUs. Distributed version uses 40 search threads, 1,202 CPUs and 176 GPUs Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

Machine learning map http://scikit-learn.

30 Machine learning map Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

The Five Tribes of Machine Learning Pedro Domingos: The Master Algorithm Tribe Foundation The main algorithm Symbolists Logic, philosophy Induction Connectionists Neuroscience Backpropagation

31 The Five Tribes of Machine Learning Pedro Domingos: The Master Algorithm Tribe Foundation The main algorithm Symbolists Logic, philosophy Induction Connectionists Neuroscience Backpropagation Evolutionaries Evolutionary biology Genetic programming Bayesians Statistics Probabilistic inference Analogizers Psychology Kernel machines NB: We ll cover one Bayesian algorithm and one Symbolists algorithm Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

32 Outline 1 Intro to Machine Learning 2 Approaches to machine learning 3 Supervised learning 4 Algo 1: Naïve Bayes Classifier 5 Algo 2: Decision Trees Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

33 Training and prediction Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

34 Training and prediction A supervised model first needs to be trained on labeled data, and after that the model can be used for prediction (classification or regresssion) on previously unseen data Each data instance is represented as a vector of features x, where each feature enodes one aspect of the input (e.g., height, weight, income,... ) Feature extractor generates the feature vector x for each data instance (this step is identical for both training and prediction) During training, the model is fed data instances x and their labels y, i.e., the pairs (x, y) After the model has been trained, the same model is used for prediction During prediction, the model is presented only data instances x as input, and it generates y as output Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

35 Overfitting Our main aim is for the model to generalize well If the model is very complex, it will fit the data very well, and make (nearly) perfect predictions If the model is too complex, it will adapt to much to the data it has been trained on, but give poor predictions on unseen data overfitting An overfitter model generalizes poorly and is therefore useless! Overfitting is the biggest problem of machine learning To avoid overfitting, we have to test how well the model will work on unseen data cross-validation Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

36 Cross-validation As unseen data are now available, we take out a portion of available data to act as unseen data We split the dataset into a training set and a test set (typically 70% 30%) We train the model on the train set, use this model to obtain predictions on the test set, and then compute the accuracy Model s accuracy on the test set indicates how well the model generalizes (the model has previously not seen the data from the test set) Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

37 Cross-validation Some models allow their complexity to be fine-tuned (e.g., prunning of decision trees) To find the optimal complexity, we need a third dataset on which we can test the accuracy of the model: the validation set We split the dataset into a training set, a validation set, and a test set, e.g.: We train models of varying complexity on the train set and test each on the validation set After we have found the best-performing model on the validation set, we compute the final accuracy of this model on the test set Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

38 Outline 1 Intro to Machine Learning 2 Approaches to machine learning 3 Supervised learning 4 Algo 1: Naïve Bayes Classifier 5 Algo 2: Decision Trees Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

39 Please refer to last year s slides. HR: NaivanBayesovKlasifikator.pdf EN: NaiveBayesClassifier.pdf Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

40 Outline 1 Intro to Machine Learning 2 Approaches to machine learning 3 Supervised learning 4 Algo 1: Naïve Bayes Classifier 5 Algo 2: Decision Trees Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

41 Please refer to these slides: UI-10-StablaOdluke.pdf (currently only in Croatian) Dalbelo Bašić, Šnajder (UNIGZ FER) AI Machine Learning AY 2016/ / 41

CSL465/603 - Machine Learning

CSL465/603 - Machine Learning Fall 2016 Narayanan C Krishnan ckn@iitrpr.ac.in Introduction CSL465/603 - Machine Learning 1 Administrative Trivia Course Structure 3-0-2 Lecture Timings Monday 9.55-10.45am